LOG IN

WHAT’S NEW

Changelog

A transparent record of the capabilities, workflow improvements, and fixes now available in ZeroGen.

Sturdier video saves and previews in the API

Finished videos survive brief provider interruptions, and ZeroGen API asset reads now include preview images and clips.

FIXES & IMPROVEMENTS

  • A finished video no longer fails when saving it from the model provider is briefly interrupted. ZeroGen retries the download over the next few minutes and reports a failure only if it keeps failing.
  • ZeroGen API asset reads and run outputs now include previews, so an app can show a grid of Library assets without downloading the originals. Each asset carries links to WebP images up to 96 and 512 pixels wide and, for a video, a short silent clip. The links work for 5 minutes without an API key, and a new output's preview appears once ZeroGen has made it.

Speech, lip sync and FLUX.3 Max, plus Library access in the API

ElevenLabs speech in a new Audio tool, sync-3 lip sync, FLUX.3 Max and keyframed video, and Library access, downloads and a Python library for the API.

HIGHLIGHTS

  • Turn a script into speech with ElevenLabs Eleven v4, in Generate's new Audio tool or by asking the Agent or a connected assistant. Pick one of 21 voices, play its sample, and direct the delivery with tags like [whispering]. You get a WAV or MP3, priced by the characters in your script.
  • Lip-sync a video or a still image to an audio clip with sync-3, in Generate's Video tool or by asking the Agent or a connected assistant. You get an MP4, priced per output frame, so a 24 fps video costs less per second than a 30 fps one. When the video and audio lengths differ, choose whether the video bounces, loops or retimes to fit the audio, the result stops at the shorter clip, or it keeps the whole video, silent past the audio.
  • FLUX.3 Max is now available in Generate, the Agent, and connected assistants. Choose Max in FLUX.3 to generate video from a prompt or animate a start frame, with an optional end frame, at 720p or 1080p for 5 to 20 seconds, with or without audio. Max follows prompts more closely and renders faster; set Reasoning to Low for a quicker render. Max has no video extension, so extending a clip stays on Standard. Through November 7, FLUX.3 Max costs a quarter of its regular price; after that, it costs the same as FLUX.3.
  • FLUX.3 and FLUX.3 Max can now guide a video through keyframes. In Generate, add keyframes between FLUX.3's Start and End frames and type the time each image should appear, up to 10 images in all, and the video passes through each one in turn. Keyframes cost the same per second as the tier's other video. The Agent, connected assistants and the API can send keyframes too.
  • API keys in the ZeroGen API private beta can now read and organize the Library: list projects and search assets with the Library's filters, read an asset and what it was made from, rename and favorite assets, file them in folders, create and manage folders and labels, and read naming presets and skills. Keys can also download the files your runs make, and the original of any other Library asset, through a private link that works for 5 minutes, and read which folders an asset is in. Give a key Library access as Read or Write in its permissions; existing keys keep their access until you edit them. Test keys work in the API sandbox project only.

FIXES & IMPROVEMENTS

  • The ZeroGen API has an official Python library, zerogen on PyPI, with a synchronous and an asynchronous client. Like the TypeScript library, it retries safely with idempotency keys, waits for generations to finish, pages through lists and uploads files in one call. The API docs' SDKs page now shows both.
  • API key permissions now use the API's name for generations: in Settings → API keys, the Generations resource is Runs, and its scopes are runs:read, runs:create and runs:cancel, matching the /v1/runs endpoints.

An asset menu with Convert, one-click reference fixes, and the API's OpenAPI description

A menu on every asset tile with Convert and Save frame, one-click fixes for references a model would refuse, an OpenAPI description and TypeScript SDK for the API, and faster results for unwatched generations.

HIGHLIGHTS

  • Every asset tile in the Library and in Generate's grid now has a menu with the same actions as its review, from favorites and downloads to Delete. Its Convert option resizes an image or saves it as PNG, JPEG or WebP. In the Library, a right-click opens the menu too. To keep one frame of a video, open it and choose Save frame; the period and comma keys step one frame at a time. Each copy is free, labeled Derived and saved beside its original, which stays as it was. In Generate, copies stay out of your history until you choose Show derived assets, and grid tiles are now the Library's size.
  • When a reference is too large or in a format the model can't use, Generate now says so on the reference with a button for the fix, such as Resize to 7302×5477 or Convert to PNG. The fix makes a copy beside the original, which stays in the Library, and puts it in the reference's place. GPT Image models now catch images over 40 megapixels before generating instead of failing at OpenAI, and Seedream 5, Nano Banana 2.1 and Ideogram 4.5 catch a reference that is too large and offer to downscale it.
  • The ZeroGen API now publishes an OpenAPI 3.1 description of every endpoint at api.zerogen.com/v1/openapi.json, generated from the same definitions the API validates requests with, and has an official TypeScript and JavaScript library, @zerogen/sdk on npm. It retries safely with idempotency keys, waits for generations to finish, pages through lists and uploads files in one call; the API docs have a new SDKs page that shows how. API errors now name a missing generation, project or asset plainly, and an unavailable model points to GET /v1/models.

FIXES & IMPROVEMENTS

  • Generations nobody is watching, including every ZeroGen API request, now finish within seconds of the model instead of up to a minute later, so the API's Prefer: wait returns finished results.
  • Generate's history opens on Created by Me again, so you start with your own generations. Choose Anyone to see everything your team has generated in the project; ZeroGen remembers that choice on this device, including after you reload the page.
  • Grok Imagine Video 1.5 and 1.5 Lite are no longer offered in Generate, the Agent, or connected assistants. Videos you made with them stay in your Library, where you can use them as references with another model. Grok Imagine Image 2 is unchanged.
  • Generate's note on how masks work now shows when you hover the Mask tile, instead of taking space under your references.
  • Screen readers now say each asset's name once in Generate's reference picker, followed by video or audio where that applies, instead of repeating the name.

The ZeroGen API, a cost check before big Agent steps, and folders in every review

The ZeroGen API and API keys for the private beta, a cost check before expensive Agent steps, video frames and resized images saved to your Library, Gemini 4 Argon, folders and one review window for assets everywhere, and Library filters by resolution and people.

HIGHLIGHTS

  • Organizations in the ZeroGen API private beta can create API keys in Settings → API keys, as a live key or a test key, and edit a key later to rename it, change its expiry or adjust its permissions. A key can expire after 1, 7, 30 or 90 days, any number of days up to a year, or never; new keys default to 90 days, and Settings warns two weeks before a key expires. Give a key All, Read only or Restricted permissions, setting each resource to None, Read or Write. Only a key's owner can add access or extend its expiry; organization admins can see every member's keys, remove access from any of them, or make them expire sooner.
  • Organizations in the ZeroGen API private beta can call api.zerogen.com/v1 with their API keys: generate images and video in one call, quote a request first, upload files or pass https URLs as inputs, hold a request open until the result is ready, and list or cancel generations. Keys can also read their own access, the organization's balance and monthly limits, and the models it can run, each with the JSON Schema its requests follow. Test keys rehearse all of it on placeholder media at no charge, including the failures, refusals and delays a live key can meet.
  • Before an Agent step whose input would cost more than $5, such as one in a very long conversation, the Agent now stops and asks, showing what the step will cost for its input and with its reply. On Claude models, input still cached from the conversation's last step counts at its lower cached price, so long conversations don't ask about steps that cost cents. Choose Continue to send the step, even past the conversation's budget; that covers the rest of the turn, and a later step asks again if its input would cost more than 25% above what you approved, or if it is over $5 on a different model, where the conversation's budget applies as usual. Or choose Summarize first to have the Agent summarize earlier messages and then continue.
  • Ask the Agent or a connected assistant to save a still from a video as a full-resolution PNG, or a smaller copy of an image, or one in PNG, JPEG or WebP, and it saves the result in your Library at no charge. Each is labeled Derived and linked to its source, which stays as it was, and the video frame is named after the video and the time it was taken from. Asking again for the same frame or copy returns the one you already made.
  • Gemini 4 Argon is now available in the Agent's model picker under Google, with Low, Medium, and High thinking. New conversations still start on Auto.
  • Generate, the Library, the Agent's results and Story's assets now review an asset in the same window, with the same details in the same order. References and Earlier references fold away (earlier ones start folded), the record behind the essentials sits under More details, and compared references are named by their own prompt with where they came from beside it. An upscaled Seedance video lists the draft it came from, with the draft's own references under Earlier references, and Compare can place those beside the output. Generate's review gains Show reference graph, opens earlier references, deletes with the Library's confirmation instead of a second click, and says when Reuse parameters can't apply a run. Story and the Agent gain favorite, labels, rename, Use as, Reuse parameters, upscale a draft, download with its metadata, delete and the reference graph, and Story's asset details gain an Open review button. Images without a stored preview now open at full size instead of showing "Preview unavailable", the Library's details offer Retry metadata when they can't load, and a failed rename and a refused Reuse parameters say what went wrong everywhere.
  • Every asset now shows the folders it's in, as clickable paths under Labels in the Library's asset details and in every asset review in the Library, Generate, the Agent and Story. Click any folder in a path to open it, use + to add the asset to another folder, or the buttons beside a folder to move the asset out of it or remove it from that folder. Move to, Add to folder and Remove from a folder now live there instead of in the actions list.
  • Filter the Library and the reference picker by one or more resolutions, such as 1080p and 4K for video or 1K, 2K and 4K for images, and by several people at once in Created by. Resolution, Created by and Labels now share one checkbox menu. The Agent can search by resolution too.

FIXES & IMPROVEMENTS

  • On GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna, Gemini 3.1 Pro (Preview) and Claude Haiku 5.5, a large message in a long conversation is now sent instead of stopping the Agent, and summarizing earlier turns always goes ahead. The Agent now reports a message as too large only when that one message is too large to send.
  • Generate's history now opens on Created by Anyone, so you see everything your team has generated in the project. Choose Me to see only your own; ZeroGen remembers your choice in each project, so switching to another project and back keeps the filter you picked.
  • The Agent and connected assistants can now upscale a teammate's Seedance 2.5 480p Draft in the same project, as the Library can. The upscale is your own generation and is billed to your organization like any other.
  • Clicking Cancel on a running generation in Generate now cancels it. Reuse parameters and Copy prompt had been sitting on top of Cancel, so the click copied the prompt instead.
  • When an upload is skipped because your Library already has the file, the upload list now names the existing file and links to the folder it is in. Click the folder to go there and open that file.
  • A reference tag in the Generate prompt that no longer points at an attached reference stays red but can now be clicked, so you can pick its replacement from the same menu as any other tag instead of retyping it.

Krea 3, Claude Haiku 5.5, and clearer billing holds

Krea 3 image generation and editing, MiniMax H3 Max at 1080P, Claude Haiku 5.5 in the Agent, clearer balance holds, and own-key cost shown at list price.

HIGHLIGHTS

  • Krea 3 is now available in Generate, the Agent, and connected assistants. Generate an image from a prompt, or add up to eight images to edit or combine them; refer to each one in your prompt by its position. Choose 1K or 2K at any aspect ratio, or leave the size on Auto to let Krea pick the shape at its largest size, about 4 megapixels. Pick Quality or Speed, and how much Krea reasons about your prompt; High reasoning, for edits only, plans the edit more carefully and keeps unedited areas of a single image. Krea 3 is priced by the exact megapixels of the output and of each image you add, so a smaller output or fewer images costs less, and High reasoning adds 5 megapixels.
  • Claude Haiku 5.5, Anthropic's fastest and lowest-cost model, is now available in the Agent's model picker under Anthropic, with Low, Medium, and High thinking. It takes up to 8 images or video posters, as its attachment menu shows, and long conversations on it are summarized sooner than on other models. New conversations still start on Auto.
  • MiniMax H3 Max now makes 1080P videos, and videos from 1 to 15 seconds long, as Turbo does, from text, frames, or references.

FIXES & IMPROVEMENTS

  • Generations that run on your organization's own provider key now show their cost at provider list price in Generate and wherever you approve them. ZeroGen still charges nothing for them. Approval modes and a connected assistant's per-generation limit now use that price, so a generation above your auto-approve price asks first, whichever key pays. A budget checks that each one fits what is left, but doesn't count own-key generations as spent yet. When there's no provider list price for a request's settings, it always asks for your approval.
  • Ask the Agent or a connected assistant about your balance and it lists the most recent holds in your projects, each with its generation, project, and amount, and whether its work is still running, has ended, or needs attention from ZeroGen. A generation's details now say whether its amount is held, charged, or released, so a held amount is never mistaken for a charge.
  • LTX 2.3 Quality and its video transforms, including Clean Plate, Render to Real, Inpaint, and Outpaint, are no longer offered in Generate, the Agent, or connected assistants. Videos you made with them stay in your Library, where you can use them as references with another model. LTX 2.5 is unchanged.

Reference Library media in the Agent

Drag Library media or Generate outputs into the Agent, and filter references by label in Generate.

HIGHLIGHTS

  • Drag images and videos from Library, or outputs from Generate, onto the Agent chat to reference them in your message, including every selected Library item at once. Type # in the Agent message box to pick a project folder and add its images and videos, newest first, up to 24 per message. Each reference shows as a thumbnail above the message, with an X to remove it, and as a named pill in the box. Uploaded files work the same way once they are saved to your Library. The Agent looks at each item's preview, or a video's poster frame, only when it needs to, so attaching media no longer uses the conversation's image allowance up front.

FIXES & IMPROVEMENTS

  • Adding a reference in Generate now has a Labels filter, as Library does. Choose one or more labels to show only assets with any of them, in All assets or a folder; Clear filters shows everything again.
  • After you open another asset from the reference graph, choose Back at the graph's top left to return to the asset you came from, one step at a time, with the graph still open. Back also stays available while the graph is full screen.
  • A 1080p video upscaled from a 480p draft now shows that draft as its source in the reference graph, and the draft shows the videos upscaled from it. Library details list the draft as Draft rather than Source, Earlier references show what the draft was made from, and downloaded and LucidLink metadata files include it in the reference history.
  • Comparisons made only of images can now be downloaded. In the tiled view, Download comparison saves one PNG with the references and the output laid out as on screen.
  • Long Agent conversations keep about four times as much history before the Agent summarizes earlier turns. When a single message is too large for the selected model, the Agent now says so instead of reporting a general error.
  • The Agent and connected assistants can now mark an image or video reference that shows a real person, as the person toggle in Generate does, on models that accept one such as Seedance, instead of having the request refused. If a generation is still refused for an unmarked real person, the assistant learns which references to mark before trying again. A reference an assistant gets wrong now says what is wrong with it, such as a missing file, instead of Invalid input.
  • Video play links from connected assistants now say the date and time they expire, seven days after they are made, instead of saying they last five minutes. Download links add the date to their expiry time.
  • When the provider refuses a video before starting it, for example because a reference appears to show a real person, the amount held for it is released at once instead of staying pending in your balance. Refused videos still on hold from before are released too.
  • Seedance video estimates for a set length and aspect ratio no longer come in slightly below the charge. They were about 1% low for a 4-second video.

Import LucidLink folders with their subfolders

Bring a whole LucidLink folder into Library, keeping its subfolders, with the same Merge and Keep both choices as folder uploads.

HIGHLIGHTS

  • Open a folder in Library's LucidLink source and choose Import with the folder's name, beside Import, to bring it in with its subfolders. ZeroGen recreates the folder inside the destination shown beside Import, or at the top level from Unfiled, then imports each image, video, and audio file into its matching folder, one at a time, with a row per file. If a folder of that name is already there, choose Merge to add only what is missing or Keep both for a second copy; files already in your Library are added to their folders, not copied. Other file types, empty files, files over 200 MB, and app or library packages are listed as skipped. Stop finishes the current file and leaves the rest for later: import the folder again and choose Merge to finish. A folder too large to import at once, such as one with more than 1,000 files to import, asks you to import its subfolders separately.

FIXES & IMPROVEMENTS

  • See any asset in the reference graph at full size without leaving the graph: choose the expand button in a tile's corner, View full size beside a selected asset, or double-click the asset in the middle. Images open in full, and videos and audio play with controls. Escape returns you to the graph.
  • Leaving Library during a LucidLink import finishes the current file and stops the remaining queue, even when navigation takes time. Reloading or closing the tab pauses the queue before the browser asks; if you stay, choose Resume import to continue.
  • Choosing a skill in the Agent message box, from the / menu or Skills, adds it to your message again instead of showing a file type error. Restore draft and reconnect brings back a failed message's attachments too.
  • Label dots on Library tiles stay visible when a tile is selected and the details panel narrows the grid. Click a tile's label dots, or the tag that appears on hover when it has none, to add, remove, or create labels right there.
  • Adding a reference in Generate now opens on the folder you're working in, so its media is ready to pick. Choose All assets or another folder to browse elsewhere.
  • Labels now sit near the top of an asset's details. In Library they follow the asset's actions; in Generate and Agent reviews they come right after the title, followed by the references.
  • Generate history has a Created by menu, like Library's. It starts on your own generations; choose Anyone to see everything made in the project, with each teammate's name beside their generations. The history entry in Generate's sidebar is now called Generations.
  • References in a creation's details are now named by the tag the prompt used for them, such as @image1, @video1, or @audio1, and those tags show as pills with a thumbnail in the Library prompt, so you can see which reference each part of the prompt means.

Draft Seedance videos with assistants

Ask the Agent or a connected assistant for a low-cost Seedance 2.5 480p Draft, then upscale the output you like to 1080p.

HIGHLIGHTS

  • The Agent and connected assistants can now make Seedance 2.5 480p Drafts, as Generate does, and upscale a draft output you choose to 1080p with the same prompt and seed within seven days. Each upscale is its own generation, quoted and approved like any other. Assistants can see when each draft output stops being upscalable and whether an upscale is already running, and they upscale only your own drafts.

FIXES & IMPROVEMENTS

  • Seedance generation cards now show whether each video was made with Reference media, Edit video, or Extend video, beside the model name, so you can compare runs in different modes without opening each one.
  • LucidLink imports now ask before you reload or close the tab while files are still importing. Leaving the page mid-import finishes the current file and stops the rest; import them again and any files already in your Library are skipped.

Let assistants generate within limits you set

Choose whether the Agent and connected assistants ask before every generation, or approve generations up to a price each, within a budget, or both, and give each project a monthly budget.

HIGHLIGHTS

  • Choose how the Agent and connected assistants get generations approved: ask every time, auto-approve generations up to a price each, allow a budget and pause when it is spent, or both. Pick a mode beside Approve, in an Agent conversation, or in Settings → Connections; it stays until you change it. Auto-approved generations show a receipt, and when a budget is reached you can add budget or switch back to asking.
  • Give a project a monthly budget under Monthly budget in Project settings: a spending limit for the project each calendar month, alongside your organization's monthly spend limit. It resets on the 1st (UTC). Organization admins and owners set, change, or remove it after confirming, and every change is recorded. Once the project's spending this month reaches its budget, new generations and Agent requests in that project stop until it resets. Project settings shows what the project has spent this month and what is left.
  • The Agent panel is now a window you can move, resize, and close. Drag it by its header to anywhere in the window, where it snaps to the edges, and drag any edge or corner to resize it, or use the arrow keys. Closing keeps your draft and any running work for the next time you open it. The Agent button appears only while the panel is closed; drag it anywhere too, and it shows how many generations in the open conversation are still running. It stays faint until you point at it. This browser remembers where you left each of them; double-click the panel's header or select Reset position and size to put the panel back. On phones both stay in their usual places.
  • Connected assistants that can run commands, such as Claude Code and Codex, can now upload images, videos, and audio from your computer to your Library, up to 5 GB per file, into a project and folder you choose. A file already in the project is reused instead of uploaded again. Assistants without that ability can still use an image or video by its web link.
  • Ask the Agent or a connected assistant to import images, videos, and audio from your organization's LucidLink filespace into a project's Library, up to 10 files at a time and 200 MB each, optionally into a folder. Each file reports whether it was imported, was already in the project, or needs another try.
  • Connected assistants can now use your skills, as the Agent does: they can list and read your personal skills, your organization's skills, and the shared skills of a project you name, then follow them. Skills you turned off stay off, and you still create and edit skills in ZeroGen.
  • The Agent and connected assistants can now do more of what you do in Library: trace what an asset was made from and what used it, see what one of your runs was built from, pin and unpin folders, and, for organization owners and admins, create, edit, archive, and restore labels. They can also check your organization's monthly spend limit and a project's monthly budget with this month's spending, and load a model's built-in prompting guide before writing its prompt.
  • The Agent can now look at your Library images and video posters, such as the outputs of a generation it ran, to check what a result shows before suggesting the next step. It looks at a small preview by default and can ask for full detail to read fine print. Connected assistants also get the faster preview by default.

FIXES & IMPROVEMENTS

  • Folder deletion choices now explain where assets go: keeping assets leaves them in their other folders or Unfiled, while deleting assets moves every active asset in the deleted folders to Recently deleted, including assets also filed elsewhere.
  • Comparing a generated video now starts beside the video it was made from, even when an image such as a start frame comes first in its references.
  • In Generate, Resolution now shows while the aspect ratio is Auto, and once you pick an aspect ratio each resolution lists the exact size it renders. GPT Image 2 and 2.5 also offer a Custom width and height, which explains any size the model can't render before you generate. The aspect ratio control shows the chosen ratio, such as 4:5, and Background has its own icon.
  • Settings → Connections now shows how to connect ZeroGen to ChatGPT on the web and in the desktop app, in place of the Codex Desktop steps. Claude, Claude Code, and Codex CLI setup is unchanged.
  • New Agent conversations on Auto now use GPT-6.1 Sol, falling back to Claude Sonnet 5.5 and then Gemini 3.8 Flash, which Auto used before. Sol and Sonnet cost more per message than Gemini 3.8 Flash; the model picker shows each model's relative cost. A conversation on Auto keeps the model it started on while that model is available, and a model you chose yourself stays selected. If your organization uses its own Anthropic or OpenAI key, Auto picks a model the Agent can run, and the picker shows that provider's models as unavailable.

Upload folders and explore how assets connect

Keep folder trees intact when uploading, browse destinations in columns, choose references with labels, and explore an asset's sources and uses in a graph.

HIGHLIGHTS

  • Drop a folder into Library, or choose Upload folder, to keep its subfolders and put images, video, and audio in their matching folders. Merge into an existing folder or keep both trees; files already in your Library are shared between them. Uploads run four files at a time and keep going when you leave the page. Stop waiting uploads and transfers, then Retry failed or stopped files while the tab stays open; files already finishing complete normally. When your browser is offline, waiting files hold until the connection returns.
  • Choose a folder for assets and run outputs with a column browser, pinned shortcuts, and search across your project's folders. Search results show their paths and open in the browser. Create a folder inside the picker, then confirm where to add, move, or import your assets. The plus beside a folder or project opens the same browser with the name field ready to type. Change the parent while naming your folder, or check Create another to keep creating folders in the same location.
  • Choose which assets a # folder adds to a prompt in Project settings. Use Favorites, or create and select labels such as hero and select to include assets with any of those labels, regardless of favorite status. Favorites remains the default, and references already in a draft stay unchanged.
  • Open the reference graph from any asset's actions to see what it was made from and what in this project used it, laid out left to right with the asset in the middle. Zoom and pan around the graph, hover an asset to see its model, prompt, and settings, and double-click one to move the graph to it; Back returns you. Videos loop when you hover them, or all at once with Play all, and full screen gives the graph the whole display.

FIXES & IMPROVEMENTS

  • Direct file uploads now support images, videos, and audio up to 5 GB per file. Larger originals can stay in your Library independently of the requirements of any generation model. URL and LucidLink transfers keep their existing limits.
  • Generation cards now have a folder icon with Add to folder and, when a source folder is available, Move to folder. Add keeps existing folder placements. Creation times are aligned to the right.
  • Project settings now has a sidebar for General, Metadata files, Reference folders, LucidLink destinations, and Skills, with only the available groups shown. Switch groups without losing unsaved changes; on smaller screens, the groups appear in a scrollable row.
  • Generate checks video and audio reference lengths against the selected model before submission, with clearer limits for Veo 3.1 Extend, Gemini Omni Flash 1.1 Edit, LTX 2.3, and LTX 2.5 Pro. LTX 2.3 Extend now offers extensions from 2 seconds and renders at 24 fps, avoiding failed 1-second requests and changes caused by the source frame rate.
  • When your browser can't play a video, the viewer now shows a notice with a short looping preview or poster when available, instead of an error dialog that blocks the close button. Download remains available wherever the viewer offers it.

Keep assets in several folders

Keep an asset in more than one folder, trace a Library asset's references back through earlier generations, and see the default values your asset names actually use.

HIGHLIGHTS

  • Keep an asset in more than one folder. Select assets and choose Add to folder to add them to another folder; they stay in the folders they're already in. While browsing a folder, Move takes your selection out of that folder and into another, and Remove from folder takes it out without deleting it; other folders keep it either way. Remove from folder, not Move, now takes assets out of a folder, so Move no longer offers Unfiled for assets. Dragging assets from a folder onto another folder moves them; dragging them from anywhere else adds them. Asset details lists every folder an asset is in, and moving a run in Generate leaves its outputs in any other folders you added them to.
  • Deleting assets while browsing a folder now asks what you want: Remove from folder keeps them in your Library and their other folders, Remove from all folders keeps them in your Library, in Unfiled, and Delete moves them to Recently deleted, where restoring them puts them back in their folders. Outside a folder, Delete also offers Remove from all folders when a selected asset is in a folder.
  • In Generate, moving a run while you're viewing a folder now takes its outputs out of that folder and into the one you choose, and Unfiled takes them out of that folder only. With no folder in view, they leave the run's current folder; if the run isn't in a folder, its outputs are only added, and Unfiled isn't offered. To add one output to a folder without moving its run, open the output and choose Add to folder; it stays in any folders it's already in.
  • A generated asset's details now show Earlier references: what its references were made from. Below the direct references, each one that was itself generated has a From group you can collapse, with earlier generations indented under the reference they came from. Select any of them to open it, or, while comparing, to place it beside the output. In the tiled comparison view, drag panes to reorder them; the numbers beside selected references follow the order on screen. Downloaded and LucidLink metadata files now include the full reference history as well.

FIXES & IMPROVEMENTS

  • The tiled comparison view now shows up to eight references beside the output, up from four.
  • Price estimates for FLUX.3 Image now match provider pricing: they follow the megapixels each size and aspect ratio renders, instead of one price per resolution. Estimates are higher at 2K and 4K, and for some wider and taller shapes at 1K.
  • Price estimates now match provider pricing for more video requests. MiniMax H3 references-to-video includes the fee for each reference image after the fifth. MiniMax H3 Max references-to-video at 480p is lower, and LTX 2.3 Quality Clean Plate, Render to Real, and Outpaint at 480p are slightly lower.
  • In the Name builder, example asset names now show each text field's default value, the text your names actually use, instead of its label. While you type a text field's label, its default value fills in to match, until you change the default value yourself. New Text fields start with Text as their default, and the Scene field in a new preset starts with Scene. Saving a new default for the preset you're using now updates that field in Generate right away, without selecting the preset again.
  • When a naming preset includes a version, each output of a multi-image or multi-video generation now takes the next version (for example v7 and v8) instead of sharing one version with _01 and _02 suffixes.
  • In Library's details panel, actions now sit right under the asset name and match the preview: Use as offers an image as a reference, start frame, or end frame, and a Seedance 2.5 draft shows Upscale.
  • Hovering a Library tile or an output in Generate's grid view now shows Use as and Reuse parameters, with Use as first wherever the two appear together. In Generate's grid view, upscale a draft from its preview or from list view.
  • After upscaling a Seedance 2.5 draft to 1080p, select Upscale again to make another 1080p version from the same draft while it remains available. Each upscale is a new generation, billed separately, and your earlier ones are kept; View opens the newest.
  • Uploading a file your project already has into a folder, or importing it there from LucidLink, now adds the existing file to that folder as well, keeping its other folders, and the notice says so. Previously it stayed only where it was. Without a destination folder, the notice names the folders the file is already in.

Find Library assets by name

Search as you type in Library, delete folders with Undo, and import LucidLink files into the folder you choose.

HIGHLIGHTS

  • Search Library assets by name as you type, alongside your existing filters. Folder searches include subfolders, and your search stays with you as you browse. Paste a file name or enter a few words to find matching assets.
  • Delete a Library folder from its menu in the sidebar, on its folder card, under Actions, or in the details panel. Choose Delete folders, keep assets to keep everything it held in your Library, where assets in no other folder show in Unfiled, or Delete folders and assets to move them to Recently deleted as well. Subfolders are always deleted with their folder. Select Undo in the notice that follows to put the folders back with their assets, pins, and LucidLink links. Deleting a folder linked to LucidLink stops its exports, and only admins can delete one.

FIXES & IMPROVEMENTS

  • When Seedance finds a real person in a reference you haven't marked as one, Generate now marks exactly that reference as a person and submits the generation again for you. The new generation replaces the failed one in your history and shows a Resubmitted label explaining which reference was marked; you can unmark it with the person toggle on the reference. When it can't resubmit, the error names only the references Seedance flagged instead of every reference.
  • When Reuse parameters can't load a run, the message now names the reason, such as settings Generate can't reproduce or a run made with an earlier version of Generate. For an upscaled 480p Draft, it suggests reusing the draft instead.
  • Library tiles for generated assets now show Reuse parameters beside Favorite when you hover over them, so you can load the prompt, references, model, and settings into Generate without opening the asset first.
  • If the folder you're generating into is deleted, Generate now says the folder isn't available and offers Clear.
  • When a connected assistant asks you to approve a generation into a folder that has since been deleted, the approval panel and review page now say so and can't be approved, instead of waiting to load the folder.
  • Library uploads now share one compact progress summary, with a bar and counter that advance together as files finish processing. Expand it to see individual files in a scrollable list. Successful uploads clear automatically after completion; errors and folder warnings stay until you dismiss them.
  • When a referenced asset is deleted from Library, Generate keeps its tag marked as no longer attached, including in restored drafts. Removing that tag keeps the other tags pointing to the same media.
  • Picking a folder in Generate now keeps its references together in one folder chip. Open the chip to remove individual references, delete it as one unit, or press Enter to edit its tags. Removing a member also detaches it when the folder pick added it and nothing else in the prompt uses it; references you added manually and frames stay. Undo restores the chip and any references its deletion removed. Your Library folder stays unchanged.
  • Pasting prompt text you cut or copied in Generate keeps each reference tag on the same media, even after the tray changes.
  • LucidLink imports now go into the folder you had open, like uploads, in Library and in Generate's reference picker. The folder is shown beside Import or Use as reference; select it to choose another folder or Unfiled. Files already in Library stay where they are, and the Import button no longer sits under the Agent button.
  • The prompt cursor no longer jumps to the start when a typed number completes a reference tag, and fast typing right after picking an @ reference or a # folder, undoing, or clicking a tag keeps its order. Pasting an image no longer later pulls focus back to the prompt or collapses selected text.

Generate and edit with Nano Banana 2.1

Google's newest Nano Banana model joins Generate, and each asset keeps one name across Library and Generate.

HIGHLIGHTS

  • Generate and edit images with Nano Banana 2.1, Google's update to Nano Banana 2 with improved image quality, prompt adherence, and text rendering. Choose 1K, 2K, or 4K and an aspect ratio, and attach up to 14 reference images to edit or combine. Nano Banana 2 remains available, including its 0.5K tier.

FIXES & IMPROVEMENTS

  • Each asset now has one name, shown the same way on Library cards, in the table, the preview, the details panel, and Generate. Library cards show the asset's name instead of its prompt, which stays in the details panel. Comparison downloads from Library's preview are named after the asset too, as they already were in Generate. Deleting the generation that made an asset no longer changes its title.
  • Library's filter bar now leads with All, Uploads and Generations, which narrow any view, including a folder. Type is a menu like Created by, Labels sit beside them, and Clear filters resets every filter from the bar. The reference picker offers the same origin, favorites, type and Created by filters. These filters replace the Uploads, Generations and Favorites sidebar entries in both places.
  • Removing a reference in Generate now leaves an empty slot in its place, and the references after it keep their tags, such as @image3, until you remove the slot. Click the slot to choose a replacement in that position, and a prompt tag that pointed to the removed reference moves to the new one. Click the slot's X to remove the slot; the next reference added in that position then takes over the tag. A reference whose usual tag is still taken by a removed reference's tag shows the next free number, and typing that tag selects it.
  • Drop a file, a generation, or one of your other references straight onto an empty reference slot in Generate to fill that slot in place, just like choosing one for it. The slot highlights while you drag over it; when you drop several files, the first fills the slot and the rest are added at the end.
  • Dropping another file onto a reference upload error in Generate now adds it as a reference, instead of opening the file in the browser tab. The References drop area is also slightly larger.
  • When Seedance rejects a generation under its content policy, the error now names what it rejected and why, for example that Image 1 may be subject to copyright restrictions, so you know which reference to replace. Previously every rejection showed the same generic message.
  • In fullscreen prompt editing, pressing Command/Ctrl+Enter when a generation can't start now shows why, instead of an earlier upload or folder message.
  • If you belong to more than one organization, signing in now asks you to choose an organization once instead of twice.

Add folder favorites to prompts

Add a folder's favorites as references in one pick, with a more compact editor.

HIGHLIGHTS

  • Type # and a folder name in Generate to add its favorites as references. Tab opens subfolders; Enter selects favorites, or opens a folder with children when it has no usable favorites. The active row shows key symbols for its available actions. The prompt inserts reference tags where supported, without a folder-name wrapper, and immediate Undo removes only references added by that pick. Picks use the newest favorites your model can accept and stay fixed until you pick the folder again.

FIXES & IMPROVEMENTS

  • Generate's naming preset is now a chip beside Name. Frames and References have focused help tooltips, and references scroll within three rows with a remaining count, Back to top, and an Add button in the heading.
  • Characters and Locations now use Library folders and favorites. The separate Character button, creation flow, and Character/Location picker sections have been removed; existing media remains in Library.
  • Press Escape in the prompt to expand or collapse fullscreen editing. An open reference menu dismisses first.
  • Pinned folders appear first in the # folder selector, with a pin icon and the same Tab and Enter controls. Folder suggestions start on the first item. Arrow-key selection stays active until the mouse moves, so a stationary cursor no longer overrides it in either prompt menu.
  • Deleting a folder pick's last prompt tag for a reference also removes that reference if the pick added it. Undo restores the tag and reference; references attached before the pick and Library assets stay intact.
  • Start and End frame cards now match the size of reference cards.
  • Empty prompts now place the cursor at the start of the input instead of after the placeholder, including in fullscreen.
  • Reference duration errors now show fractional seconds, so a clip just over the model's limit is no longer described as exactly at the limit.
  • Reuse parameters on a Seedance video in Library now loads its prompt, references, and settings into Generate, as it already did from the Generate feed, instead of saying the run can't be reused.
  • Generate folder counts now include uploads and everyone's generations, matching Library. A breakdown beside the folder path shows uploads, subfolders, and who made the generations, with a link that opens that folder in Library.
  • The bring-your-own-key estimate in Usage is now on by default whenever you have own-key Seedance usage, valued at BytePlus's published list price. Turn it off to see ZeroGen charges alone. Own-key usage on other models is counted but no longer estimated.
  • A generation no longer fails when the transfer of its finished output is briefly interrupted; the transfer is retried automatically. If it still cannot complete, the error now invites you to retry rather than contact an administrator.

Find your own and your teammates' work in Library

Filter Library by who made each asset, and select ranges from the checkboxes.

HIGHLIGHTS

  • Library now has a My generations view in the sidebar, and a Created by filter beside the media types that narrows any view to your own or a teammate's generations and uploads.
  • Generate and edit images with FLUX.3 Image, which renders detailed compositions and legible text. Choose 1K, 2K, or 4K and an aspect ratio, and attach up to 10 reference images to edit or combine.
  • Two faster, lower-cost image models join Generate: Seedream 5 Flash generates and edits at 2K with up to 10 reference images, and Recraft 4.1 Flash creates raster images from text.
  • Video adds lower-cost tiers for text and frame generation: Lite in Grok Imagine 1.5, and Turbo in MiniMax H3 Max with 480P, 768P, or 1080P output. Kling O3's 4K tier can now edit a video or follow a reference video, like Pro and Standard.
  • Edit just part of an image with a mask in Generate. Add a mask beside your image on Ideogram 4.5 or GPT Image models: transparent areas are changed (or white areas, if the mask has no transparency) and the rest is kept. The same mask works on every model that supports masks.

FIXES & IMPROVEMENTS

  • Team invitations now explain when someone is already a member or has a pending invitation, and give clearer guidance when an invitation cannot be sent.
  • Label each generated output directly in its feed card details. Organization admins can also create and apply a new label from the label picker without leaving their work.
  • Folder counts in Library and Generate include assets in subfolders while collapsed and show directly contained assets when expanded. Generate counts only generated assets. Click a folder to select and expand it; click the selected folder again to collapse it while keeping it selected. Pinned folders expand in place, independently of the main folder tree and other pins. Generate's compact folder browser remembers your expansion and selection when reopened.
  • In Library, Shift-clicking an asset's checkbox now selects every asset between it and the one you last selected, as Shift-clicking the asset already did. Select all is available before anything is selected.
  • A heart toggle beside Created by narrows any Library view, including a folder, to your favorites.
  • Reve 2.1 and GPT Image 1 are no longer available in Generate, because Reve is shutting down and GPT Image 2 and 2.5 replace GPT Image 1. Your existing images stay in Library. If either was your last-used model, Generate opens on the default image model instead.
  • Veo 3.1 now offers only the aspect ratios each mode accepts. Text-to-video offers Widescreen or Social story, and Auto remains when you start from frames or extend a video. Veo 3.1 Fast image references now offer 720p and 1080p, and image references are no longer available on Veo 3.1 Lite; use Standard or Fast.
  • MiniMax H3 Max now runs on its provider's current build of the model. Your settings carry over unchanged, but the same prompt can render differently than it did before.
  • Generate now restores a saved draft that holds many references, such as references kept while switching models, instead of retrying indefinitely.

Connect assistants and review generated media

Connect desktop assistants, approve generation requests in conversation, and review the resulting media.

HIGHLIGHTS

  • The External connections page now starts with desktop setup instructions and a prefilled Claude connector link on public ZeroGen deployments. After an approved generation finishes, connected assistants can show the resulting image or video poster; supported viewers can play the original video, and other clients receive a private play link.
  • Connected assistants that support confirmation forms can ask you to approve a generation in the conversation. Requests that need visual input review still open the ZeroGen review page.
  • Pin the folders you use most from any folder's menu. Pinned folders sit at the top of your project's folders in Library and Generate, and the reference picker now lists your project's folders and pins so you can choose a reference straight from a folder.
  • Generate and edit images with Ideogram 4.5, which is strong at legible text and typography. Choose 1K or 2K sizes and a quality tier, and attach up to four references alongside the image you edit.
  • Organization admins can turn on a bring-your-own-key estimate in Usage. It adds what your own-key generations would cost at provider rates, without ZeroGen markup, to your spend totals and charts. Your provider bills that usage, so its bill can differ.

FIXES & IMPROVEMENTS

  • Generate, Library, Settings, and Studio download much less code before they become usable. The Agent panel and video player now load when you first reach for them, so that first open can briefly show a loading indicator or an empty player.
  • Completed runs without an available output now appear as compact tiles in Generate's grid, with Move to folder still available.
  • Library opens the moment you click it, showing its layout, folders included, while your assets load, and it loads faster. The other top-bar links show a progress mark while their page loads.
  • Agent conversations using GPT or Grok models respond again instead of failing before the first reply.
  • Generate history and folders load faster in projects with many generations.
  • Project settings are now available in every organization. Members can view them; organization administrators and owners can make changes.
  • Seedance 2.5 now outputs MOV with YUV 4:4:4 chroma sampling by default; choose MP4 under Output format when you need it. Upscaling a draft to 1080p keeps the draft's output format.
  • Usage filters keep every user, model, and project available after you choose one, stay open while a filter applies, and update without reloading the page.
  • Video requests from connected assistants that don't choose a resolution now render at 720p wherever the model offers it, as Generate already does; Wan 3 previously rendered these at 1080p. When an assistant requests a Veo 3.1 video without choosing audio, the video is generated without audio, matching its estimate. Veo 3.1 Extend offers only 720p, the one resolution it renders.
  • Grok Imagine Image 2 adds a lower-cost Low quality option.
  • Price estimates now match provider pricing for more requests: GPT Image with a chosen quality and Auto size, GPT Image 2 at 2K and 4K, FLUX.2 Pro Edit reference images, larger Seedream 5 Pro images and extra references, Grok Imagine 1.5 reference images, Wan 3 reference videos, and LTX 2.3 Video to HDR. MiniMax H3 768P estimates are lower. FLUX.2 Pro Edit declines reference sets too large to estimate.

Preview Seedance drafts and reuse project files

Preview Seedance 2.5 drafts before finalizing them, reuse existing project files, and deliver metadata automatically when enabled.

HIGHLIGHTS

  • Seedance 2.5 now starts at 480p Draft, so you can review a video before choosing a separately priced 1080p upscale. Start it from the Generate card or Generate and Library review; the action shows the remaining time, and the quote uses the draft's saved duration and aspect ratio when available. When complete, View opens the 1080p result from the draft. The final result keeps the original prompt, shows the playable draft as its Source, and advances a versioned naming template.
  • Uploading or importing a file your project already has now reuses the existing file instead of adding a duplicate. The existing file keeps its name and folder, and you’ll see where it already lives. Reuse matches file contents, so edited files, and files in other projects or Recently deleted, still upload as new files.
  • When metadata files are enabled for a project, downloads and LucidLink exports of generated assets include JSON metadata automatically. Library card and bulk downloads start directly; the arrow in asset details and previews still offers media only or metadata only.

FIXES & IMPROVEMENTS

  • Open a project's settings from the menu beside that project in the project switcher.
  • Click a completed upload’s name to open the file, even when it lives in another folder.
  • Price estimates and charges for GPT Image 2.5 Sunburst and Flare now reflect the selected resolution, aspect ratio, and quality, including 2K and 4K, instead of a 1K reference price.
  • In Library, Asset details shows each asset’s ID and uses the same asset name as Generate instead of the generation prompt. Editable asset names there and on Generate cards show an edit icon on hover or keyboard focus.

Project organization and generation metadata

Keep project delivery settings organized and generation records more complete.

HIGHLIGHTS

  • Download a generation's JSON metadata file from the Library inspector or preview windows, either with its media or on its own. A single project setting enables metadata files for both browser downloads and LucidLink exports; each contains the complete available generation record, with unavailable values marked explicitly. Bulk downloads can include a metadata file for every selected generation.

FIXES & IMPROVEMENTS

  • Naming presets are shared within each project and managed by administrators. Project settings lists Library folders linked to LucidLink destinations, with actions for administrators to change or unlink them.
  • Generation metadata files distinguish requested dimensions from named resolution tiers, including image aspect ratios that Reuse parameters can reconstruct. Newly generated videos record actual dimensions, duration, frame rate, codec, and container from the delivered file, while references include their SHA-256 checksums.

Metadata files on LucidLink export and more flexible asset naming

Send a JSON metadata file beside each generation you export to LucidLink, rename assets directly from Library and Generate, and refine how naming presets build your filenames.

HIGHLIGHTS

  • Projects set up for metadata files can export a JSON file beside each media file sent to LucidLink, using the same name plus .json. Use the Include metadata files option in the export dialog to send them with the media or leave them out. Existing files are never replaced, and a later export adds any missing metadata files beside media already in the folder.

FIXES & IMPROVEMENTS

  • Files exported to LucidLink are named exactly like the asset in your Library, without an ID suffix. Exporting an asset again after renaming it creates a second file under the new name.
  • Ask your agent to export generated assets to an existing LucidLink folder, with checks that prevent overwriting files.
  • The Agent and connected external assistants can organize project folders, rename and label assets, set favorites, and use naming presets for approved generations. Download originals, including large video and audio files, directly from Agent results.
  • Repeat LucidLink exports skip files already in the destination more quickly. Choose Verify existing files to check their contents, with verified files counted separately from files already there.
  • Pinned LucidLink folders use pin icons. Destination pickers list the current directory’s unpinned folders below the pinned shortcuts.
  • Click an asset name in the Library inspector, preview, Generate card, or generation review to edit it. Press Enter or click away to save, or Escape to cancel. Original generation prompts stay unchanged.
  • New naming presets start with an empty Scene value. Switching counter sequences converts the starting value and shows only the relevant controls. Parent Folder fields can remove spaces from folder names.
  • The Library details panel stays out of the way until you click an asset or folder, giving the grid the full width. Close it from its header; clicking any asset or folder brings it back while keeping the selected card in view.

Organize project media and export to LucidLink

Send Library generations to your connected LucidLink filespace, and organize uploads and outputs across Library and Generate.

HIGHLIGHTS

  • Name generated assets with freeform text or reusable presets combining text, counters, authors, dates, and parent folders. Preview names as you edit, reorder or repeat fields, and manage shared organization presets. Optional automatic counters use the next available number in the destination folder. Naming choices restore with your draft when you refresh the same browser tab.
  • Select generations, choose a LucidLink folder, and export them together. Follow each file’s progress, stop after the current file, or retry failed exports. Existing files stay unchanged.
  • Administrators can connect a Library folder to a LucidLink destination. Members can then export its generations on demand.
  • Organize Library files into folders and subfolders. Drag files or folders into a destination folder. To move files and folders together, select them and choose Move. Subfolders appear above your files, and clicking a folder name in the path takes you back to that folder.
  • Delete assets into Recently deleted, then restore them individually or together to their original folder. If that folder is gone, restored assets appear in Unfiled. Use Move to return active assets to Unfiled without deleting them.
  • Browse your generation history by folder and file new outputs into the folder you’re viewing. Move finished runs and their outputs together by dragging or choosing Move to folder, including failed runs with no outputs. Create, rename, and move folders directly from Generate, or return to My generations.

FIXES & IMPROVEMENTS

  • Library uploads started inside a project folder file there even if you navigate away during the upload. Upload progress explains when a file uploads successfully but cannot be filed into its destination.
  • Counters added or moved to the end of a generation naming preset now increment automatically by default. You can still turn automatic increment off or explicitly start numeric versions at zero.
  • Generate cards show asset names above prompts, and Library uploads display names without file extensions. Downloads keep their full filenames.
  • Generation errors give clearer guidance for full reference-asset pools, account limits, credentials, and unsupported requests. Failures that need investigation no longer suggest an immediate retry.
  • Library selections keep bulk actions, Select all, and Clear together in the header. Export linked folders or selected generations from Actions or the LucidLink destination pill, and download selected items from Actions.
  • LucidLink destination pickers show pinned folders beside the browser with simpler controls. Click LucidLink to return to the filespace root from the Library, reference picker, or destination picker.
  • Library folder, label, and export dialogs use consistently sized action buttons that wrap on narrow screens.
  • Seedance task-mode errors explain the selected and interpreted modes, identify when an automatic retry was used, and offer concise retry advice.

Clearer generation charges and media details

Review generation charges in Usage and see clearer image details and comparison download guidance.

FIXES & IMPROVEMENTS

  • Restored video generation for organizations using their own BytePlus account.
  • Usage preserves project and member attribution for generation charges and labels unavailable accounting instead of showing it as zero spend.
  • Newly generated images show their actual dimensions in media details. Image-only comparisons explain download availability consistently without overflowing the controls.

New image models and a clearer view of usage

Generate and edit with GPT Image 2.5, see where your organization's generations and spend go, and manage your account and team from focused settings pages.

HIGHLIGHTS

  • Generate and edit images with GPT Image 2.5 Sunburst and Flare, using up to 16 references, Extra High and Max quality, and transparent PNG or WebP output.
  • Organization admins and owners can review generation activity and charges from Settings → Usage, compare model performance, see billing completeness and where spend goes, and explore up to a year of trends. Export a CSV or copy a filtered view link for teammates with access.
  • Account settings now let you update your name, request a password-reset email when signed in with a password, and review or sign out your active sessions.
  • Organization settings now open General, Team, and Billing on focused pages. In Team, find people by name or email, filter by role, browse larger teams page by page, and review pending invitations alongside current members.

FIXES & IMPROVEMENTS

  • Seedance checks reference video frame rates before generation and explains when a clip needs to be re-exported.
  • Reference images are checked against the selected model's size limit before they upload, so an oversized file is explained up front instead of failing during generation. Reusing media from an earlier generation applies the same limits.
  • New generations and completed media appear in Generate and Library without a page reload, including when you return after working in another tab.
  • Video previews and reference comparisons use more compact playback controls that fade away smoothly without darkening the picture.
  • Billing activity identifies each generation's model and media type, with pagination for reviewing older charges.
  • Video creation details identify when missing usage was billed at the original estimate, with the fallback policy shown before generation.

Create more variations and keep media close

Pin LucidLink folders, generate up to four video options from one prompt, review connected media without losing your place.

HIGHLIGHTS

  • Pin LucidLink folders for yourself; admins can also keep shared project shortcuts available in External Sources in both Library and the reference picker.
  • Library and the reference picker now share All, Favorites, and media-type views, making the same assets easier to find wherever you need them.
  • Choose one to four video outputs before generating so you can explore several results from one setup and review them together.

FIXES & IMPROVEMENTS

  • Organization admins can resend a pending team invitation when someone needs a fresh email, without changing the invitation's role.
  • Seedance 2.5 keeps your selected Reference, Edit, or Extend mode instead of switching it based on words in your prompt.
  • Creation details in Generate and Library now identify who made each generation and show its settled cost.
  • Reference previews across Generate show every attached media role, including frames, audio, sources, and masks, with consistent access to full review.
  • Media review links let you inspect related references without losing your draft or place in the feed.
  • Library card actions stay visible while an expanded Use as menu is open.
  • Generated video review metadata shows the recorded Resolution.
  • Live generation timers stay anchored to the generation's actual start when you navigate away and return.

More current models, more control over prompts and references

Create with the latest image and video families while ZeroGen follows your frames and references, then move media into the right role with less setup.

HIGHLIGHTS

  • Video adds Gemini Omni Flash 1.1, Veo 3.1, Grok Imagine 1.5, Wan 3, and MiniMax H3 Max with their available text, frame, reference, edit, and extend workflows.
  • Image adds Grok Imagine Image 2 and Reve 2.1, while Recraft 4.1 Pro can switch between raster and SVG output.
  • Text, frame, reference, and extend workflows now route automatically from the media you add; Kling O3 Create and Edit share one model with a Mode selector for the remaining ambiguous edit intent.
  • Send existing images straight to References, Start frame, or End frame from review, or drag them between the feed, reference tray, and compatible frame slots, with clear guidance when a frame setup is incomplete.
  • Select multiple assets in the reference browser and add them together, or drop several files at once anywhere on the References section. Tools with fixed per-role slots, such as LTX 2.3 Inpaint's source and mask, still take one file per slot.

FIXES & IMPROVEMENTS

  • A new reset control clears the prompt and every attached reference in one action while keeping the selected model and output settings.
  • Fullscreen prompt editing gives attached references a larger, wrapping tray with more room for writing.
  • Generate and Library now share the same media-review actions, provenance, reference comparisons, and continuous navigation across creations.
  • Generate cards and media review adapt more cleanly as the workspace narrows, and queued cards keep their final layout instead of shifting when details arrive.
  • Library delete confirmations now behave consistently across cards, bulk actions, inspectors, and previews, including reliable keyboard focus.
  • Authenticated workspace screens use consistent typography across pages, menus, and dialogs.
  • Fullscreen review focuses on the media itself while keeping provenance and details outside the viewer.
  • Seedance retries one transient task-classification mismatch automatically; failures that remain include the model's safe explanation alongside retry guidance, and completed generation failures settle promptly instead of appearing stuck in progress.
  • Generate keeps model choices aligned with verified support and safely falls back when a saved model is no longer available.
  • Generated video details use the model's reported duration when available, making Duration in Library and review more accurate.

Bring production media into one organized Library

Import media directly from a connected LucidLink filespace, then organize every asset with shared, color-coded labels.

HIGHLIGHTS

  • Browse nested folders in a connected LucidLink filespace, inspect file metadata and previews, copy several selections straight into the active project's Library, or import one previewed file into a reference setup without routing large files through your workstation.
  • Imported media arrives with useful previews and dimensions, while unsupported files and connection problems stay visible beside the affected item.
  • Admins can create, rename, recolor, archive, and restore labels, and every member can assign any combination to one file or a whole selection.
  • Select several labels to find files matching any of them across the complete Library.

FIXES & IMPROVEMENTS

  • Invited members can complete sign-in successfully whether they follow their invitation or sign in directly.
  • Hover or keyboard-focus a Generate reference to inspect it at a larger size without opening the full review view.
  • Rename Library assets directly from their titles, with quick save and cancel controls.
  • Library cards and tables keep names and media-specific metadata compact, preserve full names on hover, and use lighter selection controls.
  • When a reference does not fit the selected model, the review view now explains why instead of appearing unresponsive.

Faster ways to scan, review, and reuse creations

Browse history as a grid or list, preview motion and references in place, and carry finished work back into Generate without rebuilding your setup.

HIGHLIGHTS

  • Switch the Generate feed between a card list and a compact grid, then use the thumbnail rail to jump straight to any creation; the selected view is remembered.
  • Video cards play lightweight, muted motion previews while you browse, without loading the full-quality clip until you open it.
  • Each creation shows the references that guided it; preview them at a larger size or send an image back to Generate as a reference or video frame.
  • Reuse parameters restores a creation's prompt, references, model, and compatible settings, while Use as reference adds it to the current Image or Video setup without changing the selected tool or model.
  • Expand the prompt box into a fullscreen editor with the same draft, references, tokens, shortcuts, typeahead, character count, and clear action.

FIXES & IMPROVEMENTS

  • Output and creation navigation now stay distinct, preserve your place, and remain usable on narrow screens.
  • Step back through drilled Library previews with the in-view back control or browser Back.
  • Stored thumbnails and motion previews make visual browsing faster and more consistent across large histories.

See and reuse every creation's settings

Inspect the prompt, model, references, timing, and other settings behind a creation, then restore them in Generate.

HIGHLIGHTS

  • Feed cards and the review view show the submitted prompt, model, output count, resolution, references, identifiers, and timing.
  • Reuse parameters restores the same compatible draft directly in Generate.

FIXES & IMPROVEMENTS

  • Temporary submission failures recover more reliably instead of interrupting otherwise valid generations.

Higher-fidelity Seedance 2.5 exports

Choose 1080p or MOV output when a Seedance workflow needs more color information for finishing and compositing.

HIGHLIGHTS

  • Choose 1080p for 10-bit output, or select MOV at 480p, 720p, or 1080p for YUV 4:4:4 chroma sampling in color grading, keying, and compositing workflows.

FIXES & IMPROVEMENTS

  • Comparison exports continue while short preview clips loop.
  • Long generated prompts stay compact in the Library inspector without shortening uploaded filenames.

More reliable comparison downloads

Comparison exports explain what is missing and recover more gracefully when a browser cannot produce the preferred format.

FIXES & IMPROVEMENTS

  • Unavailable comparison downloads now explain which media prerequisite is missing, prefer MP4, and retry with WebM when the browser needs a fallback.

Switch video models without starting over

Selected media carries across video model changes, so you can explore different models without rebuilding your setup.

HIGHLIGHTS

  • Compatible images are reassigned to the right inputs, while media the new model cannot use is visibly set aside instead of being removed.

FIXES & IMPROVEMENTS

  • Earlier image generations continue resolving when newer submissions take focus.
  • Duration errors now state the minimum length required instead of returning a generic invalid-request message.
  • Copy a safe support receipt when a generation needs investigation.
  • Reference comparisons follow one synchronized timeline, and Library review controls keep prompts, actions, and keyboard focus clearer.
  • The transparent ZeroGen mark now appears cleanly in browser tabs.

Compare renders and manage Library media in place

Judge outputs against their references, then download, delete, and multi-select Library media without leaving the workspace.

HIGHLIGHTS

  • Compare image and video renders against source references with overlay, wipe, and tiled views, then export video comparisons as a single review-ready file.
  • Library cards support direct download and delete, plus Command/Ctrl-click and Shift-click multi-selection.

FIXES & IMPROVEMENTS

  • Generated assets show prompt-derived names and original prompts, uploaded audio keeps its filename visible, and downloads use readable, collision-safe filenames.
  • Reused video settings keep the correct mode and editable controls.
  • Comparison toolbars are clearer, and video remains bright when playback controls appear.

More range and control for video work

New LTX families and stronger reference handling make complex generation work easier to direct and follow.

HIGHLIGHTS

  • LTX 2.3 adds focused transforms including inpaint, outpaint, extend, clean plate, and video-to-video workflows.
  • LTX 2.5 adds Fast and Pro options for text-, frame-, and audio-guided video generation.
  • Switch tools without losing references that may become compatible again.

FIXES & IMPROVEMENTS

  • Active and concurrent generations stay visible, recover after background handoffs, and keep Copy prompt and Reuse parameters available while work is running.
  • The Generate panel can expand to half the workspace while keeping results in view.
  • Reference and upload limits match the selected model, with one consistent image-upload allowance across input surfaces.
  • Video review uses consistent controls and one shared volume setting across the tab.
  • Generation failures provide clearer, safer guidance about what to fix or retry.
  • Generate removes duplicate balance and progress callouts while preserving useful status.
  • Account recovery can switch to a different signed-in account without losing the intended destination.

More flexible references and clearer generation feedback

Keep uploaded and generated references together, direct Seedance with bracket references, and recover from setup problems with clearer guidance.

HIGHLIGHTS

  • Reference History includes uploaded media alongside generated assets.
  • Seedance 2.5 understands bracket references in the prompt for visual and audio guidance.
  • Swap and reorder attached references without breaking the prompt's links to them.

FIXES & IMPROVEMENTS

  • ZeroGen remembers the most recently selected Generate tool.
  • Portrait generations remain fully visible in feed cards.
  • Failed generations keep Copy prompt and Reuse parameters actions.
  • Validation messages explain what to fix, and estimates refresh after reference changes.
  • Generate uses clearer Character labels throughout reference setup.
  • Image uploads use the same limit across every reference surface.

More room to refine and reuse prompts

Small Generate improvements make longer prompts easier to edit and restore more compatible video setups.

FIXES & IMPROVEMENTS

  • Resize the prompt editor vertically and keep the preferred height on the same device.
  • Reuse parameters restores more compatible Seedance and Kling video generations.

A faster, more useful Library and reference workflow

Preview and organize large media collections quickly, then prepare references with model-aware guidance.

HIGHLIGHTS

  • Visual assets use stored, bounded preview stills for faster browsing in Library and Generate.
  • Assign color labels and tags to Library assets, then download any asset directly.
  • Search and preview prompt references as you type.

FIXES & IMPROVEMENTS

  • Reference prompts adapt their syntax and ordering to the selected model.
  • Image edits route to the correct workflow for the selected model family.
  • Generation estimates and automatic video controls more accurately reflect the configured mode, duration, and references.

ZeroGen launches as a model-agnostic AI production platform

ZeroGen gives creative teams one shared workspace to generate images and video with curated models, preserve the prompts and references behind every result, organize media by project, and manage review and spend together.

HIGHLIGHTS

  • Generate is the default workspace, with one-to-four image candidates and production-ready image and video workflows.
  • The launch lineup includes Seedance 2.5, FLUX.3, MiniMax H3, Kling v3, and Kling O3 for text-, frame-, and reference-driven creation.
  • Projects keep generation and Library work scoped to the production you are working on.
  • The Library shows prompts, provenance, and ordered references, and uses one consistent player across image, audio, and video review.
  • Organization balances are visible in the top bar, with self-serve credit top-ups from billing settings.
  • Deleted generations move to recoverable Recently deleted instead of disappearing immediately.

FIXES & IMPROVEMENTS

  • Mark Library assets as favorites from cards, lists, and review views.
  • Reuse generated videos and compatible Seedance creations as references without rebuilding their setup.
  • Choose an explicit video duration before generating, and see a warning when an admitted input could not be applied.
  • Valid face references from Library imports and newer assets are accepted reliably.
  • Reference chips stay deduplicated, and estimates refresh when required prompt text is added.
  • Admins can see which supported provider credentials are available for an organization.
  • Settings use clearer section names, spacing, and invitation controls, while the top bar adapts more cleanly on small screens.