v2.0.0

Cubric Studio 2.0

Cubric Vision becomes Cubric Studio, with two headlines: Flows, and an agent inside the app.

---

⚠️ Updating? Check your version first

You are on 1.5.0: update as normal, in the app or with the update zip below.

You are on anything older than 1.5.0: download the full zip below. Your projects carry over. The update zip does not work on those versions: it stops with an "unknown error" and leaves your install working as it was.

• Use rented GPUs? Update to 2.0 before 15 November 2026. That day RunPod switches off the connection every version before 2.0 uses to rent GPUs. 2.0 uses the new one.
• Your first Connect after updating sets up a fresh Pod. A Pod keeps the version it was made with, so 2.0 replaces the one an older version made instead of reusing it. The models on your network volume stay; only that first Connect takes longer.

---

Important

• Cubric Vision is now Cubric Studio. Your projects folder is renamed for you. On Windows, re-pin your shortcut to CubricStudio.exe: the old CubricVision.exe goes in 2.1.
• Mac users need macOS 14 (Sonoma) or later. Every M-series Mac can run it.
• Adding or removing a model folder now needs an engine restart. Restart engine, at the end of Settings → External Connections, waits for anything you are generating to finish.
• Inpainting no longer erases on an empty prompt. Name what you want gone ("remove the tattoo") instead. Both Klein cards.
• Krea 2 Inpaint is a lot better, and takes longer. It reads your picture as reference, so what comes back sits in the scene. Krea 2 and Krea 2 NSFW.
• NVIDIA PiD is staying. Its Model Library tile no longer says it is marked for deprecation.
• Deleting a card always asks first now. Right-click → Delete opens the same box as the Delete key, and that box offers Archive beside Delete.
• Reuse Prompt needs the source card to still be there. The app no longer keeps hidden copies of your inputs, so archive a card instead of deleting it and Reuse works as before.
• Extend and New shot are gone from video History. Make the next shot, then select both clips and Combine them, or use the new Extend Video Flow.
• "Cleanup assets…" slims a project down. Right-click a project on the landing page to clear the thumbnails and previews the app rebuilds by itself. Handy before a backup.

What's new

• Flows, the first headline of this release. A flow is a whole job in one place: it asks for what it needs a step at a time, installs its own models, and hands you a finished result. Open the Flow Library with the Flows button or TAB. Eleven to start with:
• Draw It In: draw something into a photo you already have
• Scribble: draw on a blank canvas and have it rendered
• Object Stamp: take an object out of one photo and put it in another
• Character Sheet: turn a description into a character reference sheet, with the front view's head removed for video models or kept
• Outpaint: extend a picture past its edges
• Extend Video: carry on past the last frame, sound and all
• Add Foley: give a silent clip its soundtrack
• Upscale Video: double a clip's resolution
• Voice Changer: say a line and have it come back in someone else's voice
• Song: write a song from a style, a mood and your own lyrics
• Stems: split a song into bass, drums, vocals and everything else
• Head Swap and DramaBox are sold separately as Flow packages. The Flow Library shows where to get them; drop a package on the library to install it.
• An agent inside the app, the other headline. Press Agent (or A) and ask in plain words: it picks and installs models, runs generations and Flows, and looks at what it made. It asks before anything costs money. Runs on your own DeepInfra, OpenRouter or OpenAI key, or free on Ollama.
• Routines: save a chain of steps once, run it on any cards. Ask the agent to save one ("square it, upscale it, white background"), then pick it from Routines on the selection bar. Your cards are never changed.
• See which model drives the agent best. The agent's model list shows each model's score and cost per chat. Benchmark yours and compare everyone's results at bench.cubric.studio.
• Claude, Codex and Google Antigravity can drive Cubric Studio. Ask for an image, a video or a GIF and it lands in your project; they can run your routines too. Settings → Connect an agent sets it up in one click.
• Sound and speech in the prompt box. Stable Audio 3 makes sound effects, instruments and instrumental music at the length you ask for. Chatterbox reads your text aloud in a voice you record or pick, in 23 languages.
• Audio gets its own cards: wide tiles that draw the waveform and fill as they play. Click the wave to jump there.
• Dictate instead of typing. Click the mic on the prompt box or the Agent panel, or hold Ctrl+Space. Uses your DeepInfra key, about $0.0002 a minute.
• Paid cloud models, on your own DeepInfra key. No download and no GPU: Seedream 4, 4.5 and 5.0 Pro, FLUX 2 Dev, Pro and Max, FLUX.2 Klein 9B, Nano Banana 2 Lite, 2 and Pro, Seedance 1.5 Pro and 2.0, Wan 3.0, Veo 3.1 Fast and Veo 3.1.
• Remote only skips the local engine entirely. Choose it at first launch: projects open, cloud models generate, and the tools that need the engine say so instead of failing.
• The prompt box shows what a paid run costs before you press Cue.
• Nano Banana edits take up to four reference images.
• Scribble and Draw It In can render on a cloud model. Pick FLUX.2 Klein 9B (Cloud), or Nano Banana 2 Lite on Scribble, and the Flow runs on your DeepInfra key.
• Klein 9B now asks you to accept its licence before it downloads, like MiniMax H3. Request access, paste a Hugging Face token, and the download unlocks. Free to use, and what you make is yours to sell.
• Klein inpaints properly now. What comes back sits in the light, style and perspective around your mask. Both the 4B and 9B cards.
• The FLUX.2 Klein models have better quality and fewer errors, like extra limbs.
• A new + button on the Gallery and History prompt boxes brings in a reference picture from your project or your disk. On a video in History, reference models (MiniMax H3 Reference, Seedance 2.0, Wan 3.0) work from the open clip.
• Mask a picture and give it a reference at the same time. Mask what should change, add a reference with the +, and run an Edit.
• Place puts one picture inside another. Drag a picture onto an image in History, or pick Place in the Composite tools, then move and scale it and Apply. Remove background cuts the object out first.
• Videos can be upscaled with LTX, sound and all. The Upscale tool in History doubles a clip's resolution. Short clips first: a long one can still run out of graphics memory.
• GIFs have their own workspace. Make GIF from selected cards or GIF Maker from a clip, then retime, trim and crop it, or cut the subject out of every frame.
• Big photos just work. 16K images import, thumbnail and edit, and a masked edit or Detail works on just the masked area. SVGs import too.
• Shrink big pictures by megapixels. A very large import offers to shrink to the megapixels you pick, Resize shrinks by megapixels or by a fixed factor, and the agent can do it for you.
• Sharper masks on large images. Masks work at up to 4096 pixels instead of 1536, so Grow and Shrink draw a smooth edge.
• Pick a colour straight off the picture. An eyedropper sits beside the Remove Background, Paint and Paint Adjust colours, and in Scribble and Draw It In's drawing step.
• Stack: collect several cards into one. Drop a stack on the prompt box to run the current operation on every card in it, or open it to upscale, resize or crop them all together.
• Selecting cards brings up a selection bar where the prompt box was: Stack, Compare, Combine, Make GIF, marks, Download, Archive and Delete.
• Mark cards with a dot, a square or a triangle. The heart is now a mark: click for a dot, hold to pick a shape.
• FILTER shows only images, videos or audio, or only the cards you marked. The picker behind the + button has the same filters.
• The card right-click menu is reorganised, and every entry explains itself in the status bar, including why a greyed-out one is unavailable.
• The landing page has a new crew. Each of the five mascots owns a colour, and images, video, audio and prompts wear their mascot's colour across the app. The selected model's mascot peeks over the prompt box.
• The radial menu is back. Hold TAB and pick Gallery, Models, Flows or your latest workspace.
• The video player's controls sit below the prompt box, always in reach, and volume opens from the speaker button.
• The video trim bar shows the clip's sound, so you can cut on a word or a beat instead of guessing.
• Choose which speakers or headphones Cubric Studio plays through, in Settings → Audio.
• Choose where prompts are enhanced and images described: ComfyUI (the default), Remote with your own key (DeepInfra, OpenRouter, OpenAI or any compatible address), or Ollama.
• The cogwheel by the prompt shows the LoRAs your model is running: name, strength, and a switch to mute one for the next run.
• The Model Library can show only the models your GPU can run, or the RunPod GPU you rented.
• Choosing a RunPod GPU is a proper picker now: one tile per card with its price, VRAM, stock and a Gen speed bar, fastest first.
• Get an update when you want it. Settings shows a new version at the top with a button to install it, and the startup prompt has a Don't ask again checkbox.

Fixes

• Pod runs no longer hang on "Still connecting to the remote engine" after new nodes install. The app reconnects on its own.
• A generation the engine loses now fails instead of waiting forever.
• Photos turned by the camera crop, edit and import upright.
• **History a...

v1.5.0

Cubric Vision 1.5.0

A second FLUX.2 Klein, inpainting on four more models, sharper H3 video, and faster LTX.

---

⚠️ Updating? Check your version first

The update bundle in this release patches a 1.4.4 install only.

You are on 1.4.4: use the update as normal, in the app or with the update zip below.

You are on anything older: do not use the update bundle. Download the full zip instead.
Versions before 1.5.0 carry an updater that does not check which version it is patching. It
will report success and leave the app unable to start, with your projects and settings intact
but the app itself broken.

From 1.5.0 onward the updater checks first and refuses safely, so this is the last release
where the choice is yours to get right.

On a Windows build older than 1.3.0: you cannot be reached by an update at all. Smart App
Control blocks the update scripts. Download the full zip.

---

Important

• The video quality settings have been renamed, and a saved project may open at a different
  size. MiniMax H3's sizes moved down one step: what used to be the top ordinary setting is
  now the middle one, and a genuinely larger option sits above it. The sizes themselves did
  not change, the labels did, because the old top setting produced draft-grade video and
  should not have been called high. Reopening an older project may therefore show a different
  size than the one you saved.
• MiniMax H3 uses noticeably less memory than before, which makes it far more reliable on
  smaller machines and on rented GPUs, where running short of memory would previously stop a
  video partway through without explaining why.
• Rented GPUs are now chosen with enough memory to actually run H3. Renting a remote GPU used
  to place you on whatever machine was free, including ones with too little system memory for
  a video model, where a generation could be killed partway through with no explanation. New
  Pods now ask for at least 56 GB. You can change that figure in settings, and should raise it
  for anything heavier than H3.

What's new

• A second FLUX.2 Klein: the 9B. A larger, stronger sibling of the Klein already in the
  app, with its own seven styles: Storybook, Comic, Anime, Chibi, Doodle, Vintage and
  Watercolour. They are deliberately not the same eight the smaller Klein offers: four of them
  are a different artist's work entirely, so naming them after the 4B styles they sit beside
  would be a label that lies about what loads. Both cards now carry their size in the name, so
  it is clear which one you are picking and which styles belong to it.
• Inpainting comes to SDXL, Illustrious, Pony and Krea 2. Paint a mask, describe what
  should be there, and only that part is rebuilt. The rest of the picture is left exactly as
  it was. It was previously available on FLUX.2 Klein alone. On Krea 2 the fast/quality toggle
  works here too, so a quick fix stays quick.
• Sharper 2K and 4K video from MiniMax H3. H3 now works in two stages: it lays the video
  down first, then a new upscaling stage rebuilds detail as it enlarges. High resolution clips
  hold together where before they were sampled straight out at full size. Turbo also moves to
  a stronger, better-trained fast model, so quick drafts look closer to the finished thing
  instead of glossy and plastic.
• The largest file H3 needs now comes from our own servers. It used to be fetched from
  HuggingFace, where it trickled in at well under 1 MB/s and could run for hours or give up
  before finishing. It is the same file, it just arrives at a usable speed now. If our copy is
  ever unreachable the app falls back to the original source on its own.
• Longer LTX clips render faster. LTX now uses ComfyUI's own built-in fast attention,
  which needs nothing installed and works on the graphics card you already have. The gain
  grows with clip length: nothing measurable on a 2 second clip, around 13% off a 5 second one.

Fixes

• Uninstalling a model left some of its files on disk while reporting success. Anything the
  model kept next to the engine, frame interpolation weights for instance, could not be
  removed at all, and the app told you the files were being kept deliberately. They are
  properly removed now, and the space comes back.
• A half-installed model can give its disk space back. If a model lost one of its files, you
  tidied the models folder by hand, or an install stopped near the end, the library showed it
  as not installed and offered no way to remove what was still there, so the bytes were
  stranded.
• Connecting to a rented GPU no longer announces every model already on it. A connect quietly
  repairs the remote engine, and that repair was being reported as though each model had just
  been downloaded, a stack of toasts plus one naming a raw internal job id.
• Around ten console windows no longer flash open behind the app on startup, and the same
  flash is gone from video trimming, cropping, reversing and GIF export.
• Media you drag in is imported from disk instead of being copied through memory first, which
  stops a large video from stalling or failing outright. A rotated video also reports the size
  you actually see rather than its pre-rotation one.
• Applying an update leaves the install describing itself correctly, instead of reporting the
  version it replaced.
• Video crop and reverse name their output like every other tool, and the prompt box files its
  result under the operation you picked.
• Two model-library defects: the Frame Interpolation download showed another pack's name, and
  a model whose files sit outside the usual folder could read as installed to the installer
  and not-installed to the library, badge stuck, Install downloading nothing.
• The gallery no longer keeps playing a hovered clip underneath an overlay that just opened,
  and a canvas hidden behind another panel is no longer mistaken for a resize.
• Videos in the gallery looked softer than the images sitting next to them. Their preview
  frames were being generated at half the width of image thumbnails. Both are the same size now.
• The H3 quality picker offered a size the model cannot actually produce, so picking the
  highest setting could fail or hand back something other than the size on the label. The
  sizes offered now match what H3 renders.
• The video preview flashed a few scrambled frames at the start of an LTX generation.
• Downloading models to a rented GPU has had three failures fixed: it no longer gives up when
  the cheap CPU machines it prefers are all taken, it no longer starts more downloads at once
  than the machine can carry, and a download now picks itself back up if the remote machine
  restarts underneath it instead of stalling silently.

Engine

• The generation engine moves from ComfyUI 0.31.0 to 0.34.0. Rented GPUs move with it: a
  remote machine now runs the same engine version your own does, which is what the new H3 and
  LTX video work needs to run there at all.

---

First launch

The builds are not code-signed. There is no certificate and no installer. Signing would
start SmartScreen's reputation clock rather than skip it, so the warnings below are expected.

Windows. Extract the zip anywhere and run CubricVision.exe. Windows may show
"Windows protected your PC". Click More info, then Run anyway. The app is unsigned,
and this warning appears until the build earns reputation.

macOS. Any downloaded build is quarantined. Clear it, then launch:

xattr -dr com.apple.quarantine "<extracted folder>"

then double-click start.command.

---

Platform support, please read

• Windows - tested on the maintainer's Windows development machine, including a full
  in-place update from a published 1.4.4 install followed by a real generation. Not yet
  validated on a separate clean Windows host.
• Linux - not tested on this version.
• macOS - not tested on this version.

v1.4.4

Cubric Vision 1.4.4

A patch release. Three fixes, no new models, no engine change.

This is the follow-up to 1.4.3. That release gave a half-installed engine a voice and a Repair engine button — and the person who reported the problem in #2 took the update and could reach neither. This release fixes why his install kept failing, and the two reasons the repair was not where 1.4.3 said it would be.

Windows, macOS (Apple Silicon) and Linux portable builds below. Extract and run — no installer.

Updating? If you are on 1.3.0 or later, your app checks for updates on startup and will offer this version in-app on all three platforms. If you are on 1.2.0 or earlier on Windows, download the full zip below — the updater in those builds is a .bat and Smart App Control blocks it outright, so an in-place update cannot reach you. Your projects live in Documents and are untouched either way.

The engine does not change in this release (ComfyUI 0.31.0, same as 1.4.2 and 1.4.3). Nothing re-downloads.

Fixes

• The engine can now finish installing even when a shared Windows folder outside Cubric is damaged. The step that installs the rest of the engine used a working folder in your Windows user profile that every Python program on the machine shares. If a single file in there had become unreadable, the install stopped at that point — on every launch, indefinitely, with nothing you could do about it from inside the app. Cubric now keeps that working folder inside its own engine folder, where Repair engine can reach it.

• Engine health no longer appears when nothing is wrong. In 1.4.3 the section was meant to show only on a broken install, but a display rule overrode that and showed "Part of the engine did not install" to everyone — with no button beneath it, because on a healthy engine there was nothing to repair. It now appears only when there is genuinely something to fix.

• Repair engine is now offered without starting the engine first. The app only knew the engine was broken if it had tried to start it since you last opened Cubric, so going straight to Settings after launching showed nothing to repair on an install that was definitely broken. A failed install is now remembered between launches, and clears itself the moment the engine installs properly.

What's new

None. This release is fixes only.

---

Platform support - please read

• Windows - tested on the maintainer's Windows development machine. Not yet validated
  on a separate clean Windows host.
• Linux - the portable extracts and launches, and engine setup runs, verified on the
  maintainer's test laptop. That machine cannot run the engine itself, so no image or
  video has ever been generated on Linux here. Generation support is therefore unvalidated -
  reports from a capable Linux host are very welcome.
• macOS - artifacts are produced mechanically and are maintainer-untested. Community
  validation is needed before any stronger macOS support language is used.

v1.4.3

Cubric Vision 1.4.3

A patch release. Three fixes, no new models, no engine change.

All three are the same problem seen from different sides: a piece of the engine could fail to install without anything saying so, and the app had no way to tell you, no way to repair it, and no way to get the log to us. That is what #2 turned out to be.

Windows, macOS (Apple Silicon) and Linux portable builds below. Extract and run — no installer.

Updating? If you are on 1.3.0 or later, your app checks for updates on startup and will offer this version in-app on all three platforms. If you are on 1.2.0 or earlier on Windows, download the full zip below — the updater in those builds is a .bat and Smart App Control blocks it outright, so an in-place update cannot reach you. Your projects live in Documents and are untouched either way.

The engine does not change in this release (ComfyUI 0.31.0, same as 1.4.2). Nothing re-downloads.

Fixes

• A half-installed engine now says so, instead of letting a generation fail on a technical error. Part of the engine is downloaded separately after install, and if that step failed the app started anyway and looked perfectly healthy — until you picked one of the affected models and the generation died with a message naming an internal component. It now tells you plainly that part of the engine did not install, and stops those generations up front rather than halfway through.

• You can repair the engine yourself, from Settings. A new Engine health section appears in Settings only while something is actually wrong, with a Repair engine button that reinstalls the missing part and restarts the engine. It takes a few minutes and your models are left alone. There was previously no way to trigger the reinstall at all — reopening the app did not do it.

• Sending us the log actually works now. The Report on GitHub button quietly did nothing in a released build: it needed a credential no release ships with, so the click just failed in the background with nothing shown. The error window now has a Show log file button that opens the log in your file manager, and Report on GitHub opens a pre-filled report form in your browser for you to attach it to.

What's new

None. This release is fixes only.

---

Platform support - please read

• Windows - tested on the maintainer's Windows development machine. Not yet validated
  on a separate clean Windows host.
• Linux - the portable extracts and launches, and engine setup runs, verified on the
  maintainer's test laptop. That machine cannot run the engine itself, so no image or
  video has ever been generated on Linux here. Generation support is therefore unvalidated -
  reports from a capable Linux host are very welcome.
• macOS - artifacts are produced mechanically and are maintainer-untested. Community
  validation is needed before any stronger macOS support language is used.

v1.4.2

Cubric Vision 1.4.2

A patch release. Fixes and small improvements on top of 1.4.1 — no new models, no engine change.

The headline is a shape: video gets a 21:9 cinematic ratio, on every video model and at every quality level, so the wide anamorphic frame is something you choose up front instead of cropping to afterwards.

Alongside it, notes stop being plain text — Project notes and Card notes render your markdown — and an annotated card now says so on its own thumbnail. Fill finally works on the paint layer, which is the one thing an outline tool exists for.

Windows, macOS (Apple Silicon) and Linux portable builds below. Extract and run — no installer.

Updating? If you are on 1.3.0 or later, your app checks for updates on startup and will offer this version in-app on all three platforms. If you are on 1.2.0 or earlier on Windows, download the full zip below — the updater in those builds is a .bat and Smart App Control blocks it outright, so an in-place update cannot reach you. Your projects live in Documents and are untouched either way.

The engine does not change in this release (ComfyUI 0.31.0, same as 1.4.1). Nothing re-downloads.

What's new

• Video has a 21:9 cinematic ratio. It sits beside 1:1, 9:16 and 16:9 on every video model and at every quality level, so the wide anamorphic shape is a choice up front rather than a crop afterwards.

• Notes read as notes now — Project notes and Card notes render your markdown instead of showing the raw text. Headings, tables, lists, quotes and code all come out formatted, with each heading level in its own colour so a long document is easy to scan. A pencil/eye pair at the top right switches between editing the source and reading it, and the preview shows what you have typed before you save it. Long notes scroll — before, anything past the height of the window was simply cut off with no way to reach it. The window is wider and the text larger, so a table has room to breathe. Release notes take markdown too, so this "What's new" can be formatted from here on.

• A card with notes says so. An annotated gallery card now carries a small notes icon in its top-right corner, between the heart and the reuse button, and it stays visible without hovering — so a project you have been given can be scanned for the cards someone wrote about instead of opening each one to check. Clicking it opens those notes straight away.

• Fill works on paint now. The Fill button in Adjust closed enclosed holes in a mask but did nothing to the paint layer, so a closed shape drawn with the paint brush — the one thing an outline tool exists for — could not be filled in. Pressing Fill on paint now floods every enclosed area in your current colour, and every brushstroke you already made keeps its own colour on top of it.

Fixes

• MiniMax H3 video looks better with Turbo off. The slower setting was coming out worse than the fast one, which is backwards — it now gives the quality it asks the extra time for.

• Reference images and audio no longer linger in the prompt box after you move to a model or operation that cannot use them. Reusing a text-to-image card while references were staged left them sitting in the tray, and the next generation quietly ran carrying inputs it should never have had.

• Removed gallery cards no longer leave their video preview running in the background.

• Losing a cloud engine mid-generation now says the engine dropped, instead of blaming the size of your inputs.

• MiniMax H3 at 2K and 4K no longer shows LTX 2.3's motion caveat. It now reads Experimental — High VRAM.

---

Platform support - please read

• Windows - tested on the maintainer's Windows development machine. Not yet validated
  on a separate clean Windows host.
• Linux - the portable extracts and launches, and engine setup runs, verified on the
  maintainer's test laptop. That machine cannot run the engine itself, so no image or
  video has ever been generated on Linux here. Generation support is therefore unvalidated -
  reports from a capable Linux host are very welcome.
• macOS - artifacts are produced mechanically and are maintainer-untested. Community
  validation is needed before any stronger macOS support language is used.

v1.4.1

Cubric Vision 1.4.1

A patch release. Fixes and small improvements on top of 1.4.0 — no new models, no engine change.

Most of it came out of one real incident. A "No GPU — download only" cloud Pod was connected while a generation was running locally, and the app treated that Pod as its generation engine: it tried to stage model weights onto a machine with a few gigabytes of RAM, the Pod was killed, and the app read that as a dead engine and dropped cloud mode — taking four in-flight installs down with it. Six separate faults came out of that one thread, and each is fixed here.

The rest is notifications: ones that arrived long after the fact, and ones that were gone before they could be read.

Windows, macOS (Apple Silicon) and Linux portable builds below. Extract and run — no installer.

Updating? If you are on 1.3.0 or later, your app checks for updates on startup and will offer this version in-app on all three platforms. If you are on 1.2.0 or earlier on Windows, download the full zip below — the updater in those builds is a .bat and Smart App Control blocks it outright, so an in-place update cannot reach you. Your projects live in Documents and are untouched either way.

The engine does not change in this release (ComfyUI 0.31.0, same as 1.4.0). Nothing re-downloads.

What's new

• First run asks how you want to work before it asks for a folder. Setup opened on a models-folder picker, with "run on a cloud GPU instead" as a small link at the bottom — so the only obvious path was a multi-gigabyte install, even on a machine that will never use it. It now opens on the actual question: Local + Remote installs ComfyUI here and generates on your own card, with a cloud GPU there for the jobs it cannot take; Remote only skips the install entirely and takes you straight into the app to connect a RunPod account. Each side says plainly what it costs you — a one-time install and a capable GPU, or an account billed while a Pod runs — and you can change your mind later in Settings. The folder picker is still there, one step further in, and it now reads as one screen rather than three different left edges.

Fixes

• Video previews play at the video's speed again. On some runs the live preview flashed the whole clip in an instant each time the sampler advanced, then froze until the next step. It now plays continuously at the clip's own frame rate, and no longer trims the start of the clip.

• A cloud Pod used for downloads is no longer treated as a generation engine. Connecting a "No GPU — download only" Pod while you were working locally could take your generation with it: the app tried to stage model weights onto a machine with no GPU and a few gigabytes of RAM, the Pod was killed, and the app read that as a dead engine and dropped remote mode — taking every in-flight install down with it. A download Pod is now only ever a download target.

• A cloud install that dies now says so, instead of freezing. When the Pod went away mid-install, the model card kept showing the Pod's last progress — a live-looking bar and a Cancel button for a download that could never finish, sometimes painted over a model your own disk already had. Those installs now end properly, say why, and offer a working Retry, and the Model Library goes back to showing what is really on your machine.

• Losing the Pod mid-install is a notice, not an error report. Stopping your own Pod raised the "Download Failed" dialog with a Report on GitHub button. It is now a normal heads-up message.

• A queued install can no longer land on the wrong machine. Installs run one at a time, and the engine was only decided when each one's turn actually came — so a model queued for your cloud Pod could start downloading to your local disk after the Pod disconnected, tens of gigabytes you never asked for. A queued install now remembers where it was meant to go and is cancelled, with a message, if that engine is gone.

• "Generation complete" notifications no longer arrive long after the fact. A batch that did not finish cleanly could leave its finished-count behind, and the next time anything drained the queue that stale count fired — announcing generations that had ended ten or twenty minutes earlier, sometimes over an hour. Counts that outlive their batch are now discarded.

• Notifications stay up long enough to read. Every notification got the same three seconds, whether it said "Copied" or quoted three figures and a drive letter — so the longer and more useful ones were gone before the first line landed. A notification now stays for as long as its message takes to read. And when two are on screen at once, only the first counts down: the second waits its turn instead of expiring while you are still reading the one above it.

• "Not enough disk space" now names the disk, and asks for the right amount. A local install could refuse with "29.3 GB needed, 29.4 GB free" — more free space than it said it wanted — because the check quietly reserves a small working margin the message never mentioned. And when the full disk was your cloud Pod's volume, the app replaced the server's exact figures with "free up space and try again", pointing you at your own machine, where clearing space would not have helped. Both messages now name the drive or the Pod volume and quote the amount actually required.

---

Platform support - please read

• Windows - tested on the maintainer's Windows development machine. Not yet validated
  on a separate clean Windows host.
• Linux - the portable extracts and launches, and engine setup runs, verified on the
  maintainer's test laptop. That machine cannot run the engine itself, so no image or
  video has ever been generated on Linux here. Generation support is therefore unvalidated -
  reports from a capable Linux host are very welcome.
• macOS - artifacts are produced mechanically and are maintainer-untested. Community
  validation is needed before any stronger macOS support language is used.

v1.4.0

Cubric Vision 1.4.0

The biggest release since 1.0.

Minimax H3

The beast has arrived 🎉
We did everything we could to make it fast, high quality, and simply amazing.

The whole canvas toolkit arrives - mask tools, Paint,
Composite, Adjust, ten brushes and shapes - along with a new Control operation, localised
editing on four models, video that comes with its own sound, and a rewritten engine setup.

Windows, macOS (Apple Silicon) and Linux portable builds below. Extract and run - no installer.

Updating? If you are on 1.3.0 or later, your app checks for updates on startup and will
offer this version in-app on all three platforms. If you are on 1.2.0 or earlier on
Windows, download the full zip below - the updater in those builds is a .bat and Smart
App Control blocks it outright, so an in-place update cannot reach you. Your projects live
in Documents and are untouched either way.

The engine updates with this release (ComfyUI 0.31.0). It is a small update to install:
the engine moves, but nothing in the heavy machine-learning runtime does, so there is no
multi-gigabyte re-download.

Important changes

• Krea 2 and Chroma keep your source dimensions. On Krea 2, every operation except Text to Image and Image to Image now follows the input image's shape instead of the ratio picker. On Chroma it applies to Control, Detail and Upscale.

• LTX 2.3's Balanced tier moved to a better, smaller model file. The two GPU-specific options (RTX 50 Series / RTX 40 & Older) are gone — one file now runs on every card, at 20GB instead of 22–23.5GB, with better detail and sound. If you already have LTX Balanced installed, the new file downloads once and the old one is cleaned up. Text to Video and Image to Video also share one workflow now, so which one you get follows from whether you supply a first or last frame.

• Wan 2.2 is image-to-video only now. Its Text to Video operation has been retired: LTX 2.3 does that job better, and Wan's text-to-video weights were a 27GB download that served nothing else. Image to Video is untouched — same model, same speed, same results. If you already downloaded the text-to-video weights, they are no longer counted as part of Wan 2.2, so the next time you uninstall a model they get cleared out with it rather than sitting on your drive owned by nothing. Text to video lives on LTX 2.3, MiniMax H3 and Wan 2.2 5B.

• Some models now ask you to accept their licence before they download. A few model licences require us to show you the restrictions and record that you accepted them before you receive the weights — MiniMax H3 is the first one we ship. You will see the terms, and where a model is not licensed in your region, a link to request your own authorization from the licensor. Every other model installs exactly as before, with no extra step. The licence and a way to report misuse stay reachable from the model's panel afterwards.

• Your mask reaches the model the way you drew it. A masked edit used to close every hole in your mask and grow it slightly before generating, so a mask drawn as a ring around a subject arrived at the model as a solid blob — the shape you wanted was impossible to ask for. That no longer happens on any model. If you do want a hole closed, the new Fill button does it where you can see it.

• Mask composite follows the mask you drew. Compositing through a ring or an edge-band mask filled in the middle before blending, so the whole enclosed area came through instead of just the band you selected — the same thing that used to happen inside the generator and was fixed there in this version. It now uses the mask as drawn. Use Fill if you want a hole closed.

• The Add / Subtract composite dialog is gone. Blending two images used to mean selecting both, having one of them already carry a mask, and answering Add or Subtract in a dialog — with nothing on screen to judge it by. The Composite tools replace it entirely: the selected image is always the one on top, and you see the blend while you make it. Two images of different sizes are now cropped to fill rather than stretched, and the preview on the canvas matches the file that gets written.

• A detection you don't add is a preview. Detect, Points and Text show their result in green as a preview, and it stays a preview: leaving the tool discards it, and generating without pressing Add no longer sends it. Previously an un-added detection was silently used in the next generation. Because of this, the operations that need a mask stay unavailable until you press Add or Subtract — a green preview on its own is not a mask yet.

• Krea 2 and Chroma styles now reach Detail and Upscale. Every operation on these models runs one workflow, so the style you picked stays applied through the finishing passes instead of being dropped by them.

• Krea 2 Control takes a second image. Image 1 supplies the pose and composition, image 2 supplies who is posed into it.

• On a Mac, install Apple's Command Line Tools before your first setup. Setting up the local engine needs git, which on macOS only arrives with the Xcode Command Line Tools — and a clean Mac does not have them. The first Install press stops with a message naming exactly that. Open Terminal, run xcode-select --install, let Apple's download finish, then press Retry. This is a known issue, not a fault in your download: Windows ships a prebuilt engine and never needs git, while macOS and Linux build theirs. Removing the requirement is being worked on separately.

What's new

• Animate towards an image, not just away from one. Drop a single image into the prompt box on a video model and it has always become the first frame. Now the label along the bottom of the thumbnail is a button: click it and it reads Last frame, so the clip ends on that image instead of starting from it — the model invents everything leading up to it. Drop two images and you get both ends, first and last, in strip order. Useful for closing a loop, landing on a specific pose, or joining two shots you already have. Both LTX 2.3 and Wan 2.2 take it, including the end-frame-only case where the single image you give is the one the clip finishes on.

• The mask tools, rebuilt as a toolkit. Everything that makes a mask now sits in one group and works the same way. Paint one by hand with ten brushes — hard and soft rounds, feather, airbrush, chisel, calligraphy, spray, charcoal, stipple and a dry brush — instead of the single hard circle. Drop a rectangle, triangle or ellipse on the canvas and Add or Subtract it: drag the shape or its handles to size it, hold Shift to resize without squashing it, Alt over a handle to spin it, and it stays where you put it so stamping the same ellipse three times is three drags. Adjust the result — grow the whole mask, shrink it, or reduce it to a band around its edge — with a live preview and Apply, so you can fix the edge of a subject without regenerating the middle of it. Fill closes enclosed holes in one press, where you can see what is being filled before you commit to it. And a finished mask can be turned into paint from the right-click menu, or paint turned back into a mask, as a copy that leaves the original alone. No model runs and nothing downloads for any of it: all instant, all undoable.

• Detection: one button, one pass, and a way to stop it. The three detection methods — Points, Text and Auto — now open in a small strip beside a single Detect button instead of taking three toolbar slots. One detection finds every match in a single pass, so picking a result, or changing your mind and picking another, no longer re-runs it — trying all of them costs what one used to. While it works, the status bar reports it with a timer and Detect turns into Stop: a slow pass over a busy image used to look like the app had frozen, with no way to call it off.

• Localised edit — edit part of an image at full size. Mask an area before an edit and only that area is regenerated. A whole-image edit has to shrink your picture to the size the model works at; a masked edit crops to the mask instead, so the picture keeps its original resolution and everything outside the mask is untouched. It rewards a precise mask — for broad changes like a new pose, leave the mask off and let the model rework the whole image. On Krea 2, FLUX.2 Klein, Qwen Image Edit and Boogu Image Edit.

• Control — copy the structure of an image. One operation that takes a picture, reads the structure out of it and paints your prompt into that shape. A Control Type picker chooses WHICH structure: Depth keeps volume and framing, Pose keeps only the body skeleton, Scribble and Canny keep outlines. SDXL Realistic, SDXL NSFW, ILL Anime Beauty, ILL Anime and PONY Mix offer all four; Qwen Image Edit offers Depth and Pose. It replaces the old Depth operation everywhere: Krea 2 and FLUX.2 Klein still do depth, they just do it under the new name and show no picker, having only the one.

• Control on Chroma. Both Chroma models can now follow the depth and composition of an input image and paint your prompt into that shape.

• Four styles on Chroma. Chroma gains a style rack — B&W Sketch, Lenovo, Brushwork and Anime — available on every operation, Detail and Upscale included.

• Control Strength. Krea 2, FLUX.2 Klein, Chroma and the SDXL family gained a slider on Control for how hard the control map pulls. At full strength the composition is locked to the source; easing it off lets the model reinterpret the framing while keeping the pose. Klein bites softer than Krea 2, so it wants a lower setting.

• Composite — blend two of your images and watch it happen. Two new tools in a Composite group. Right-click the image you want underneath and choose Send to Composite, then open Paint Comp on another image and erase — the picture underneath shows through live where you erase, and the brush paints the top one back. Mask Comp does the same cut in one step using...

v1.3.1

Cubric Vision 1.3.1

A bug-fix release. Everything in 1.3.0, plus fixes for a class of failure that made
the app look broken when it was the network blocking it.

Windows, macOS (Apple Silicon) and Linux portable builds below. Extract and run — no installer.

Updating? If you are on 1.3.0, your app checks for updates on startup and will offer
this version in-app — on all three platforms, Windows included. If you are on 1.2.0 or
earlier on Windows, download the full zip below: the updater in those builds is a .bat
and Smart App Control blocks it outright, so updating in place cannot reach you. Your
projects live in Documents and are untouched either way.

Fixed

• YOUR ENGINE NOW KEEPS THE PARTS THAT DOWNLOADED. If anything failed while setting
  up the engine, everything that had already downloaded was thrown away — including a
  complete set of components that came down perfectly from a different server. The next
  launch saw them missing, downloaded them again, and threw them away again. For anyone
  whose provider blocked our model server, that loop never ended: the engine could never
  finish installing a single component, so generating always failed no matter which models
  were on disk. Whatever lands now stays on disk.

• A FAILED SETUP NO LONGER LOCKS YOU OUT OF THE APP. If the dependency repair on
  startup failed, the only button was Retry — and on a blocked connection Retry failed the
  same way every time, with no way past the screen. Setup now tells the difference between
  a missing component and a missing model file: if the components are all there, the app
  opens. If they genuinely are not, the error screen now has a Continue without them
  button so you can get in, and go to Settings when you are ready.

• A BLOCKED DOWNLOAD NOW EXPLAINS ITSELF INSTEAD OF LOOKING LIKE A CRASH. When
  something between you and our server interferes with the connection — an ISP filter, a
  parental or network filter, antivirus web protection, a VPN, hijacked DNS — the app used
  to show a wall of raw security-library text and offer to file a bug report, which reads
  as "the app is broken". It now names the server it could not reach, says plainly that
  something on the network is blocking it, and tells you what actually helps: a VPN for the
  whole download, with a note that changing your DNS alone is often not enough — the user
  this release was built for tested a DNS switch and his download still died a fifth of the
  way through. Real failures — a corrupted file, a full disk, a permissions problem — still
  report as themselves.

• RETRYING A DOWNLOAD NO LONGER RUNS THE PROGRESS BAR BACKWARDS. A retry resumed from
  what was already on disk but reset the bar to zero, so it looked like the whole download
  had been lost.

• AN ENGINE SETUP FAILURE NOW SAYS WHY IT FAILED, instead of listing only which pieces
  failed and dropping the reason.

• THE APP NO LONGER DIES AT STARTUP IF YOUR DOCUMENTS FOLDER CANNOT BE FOUND. It closed
  before any window appeared. It now carries on.

First launch

Both desktop platforms show a security prompt on first launch because these builds are
not code-signed. There is no certificate and no installer — this is expected.

Windows. Extract the zip anywhere and run CubricVision.exe. Windows may show
"Windows protected your PC" — click More info, then Run anyway.

macOS. Any downloaded build is quarantined. In Terminal:

xattr -dr com.apple.quarantine "<extracted folder>"

then double-click start.command.

Platform testing

• Windows — tested on the maintainer's Windows development machine. Not yet validated
  on a separate clean Windows host.
• Linux — install and launch validation only, on a low-powered Ubuntu laptop.
  Generation support is unvalidated.
• macOS — artifacts are produced mechanically and are maintainer-untested. Community
  validation is welcome.

Reporting a result? Please include your platform and OS version, CPU architecture, GPU and
driver, the artifact name, whether it was a clean install or an update, what the launcher
did, what engine setup did, whether you tested generating, and the tail of your app log.

Known limitation

If your internet provider is interfering with the connection to our model server, this
release makes the failure readable and stops it destroying your engine setup — but it
cannot get the files past the block. A VPN for the duration of the download does. A second
download location is built but not yet switched on.

v1.3.0

Cubric Vision 1.3.0

Windows, macOS (Apple Silicon) and Linux portable builds below. Extract and run — no installer.

Updating from an earlier version? On macOS and Linux your app checks for updates on startup and will offer this version in-app. On Windows, download the full zip below — the in-app updater in 1.2.0 and earlier does not work, so this release ships no Windows update bundle. Your projects live in Documents and are untouched either way.

Important

• WINDOWS NOW STARTS FROM CubricVision.exe — start.vbs and start-with-terminal.bat are gone; double-click CubricVision.exe in the folder instead. On a clean Windows 11 machine, Smart App Control blocks .vbs and .bat files outright with no way to allow them, so every way we shipped to start the app was blocked and double-clicking did nothing at all. An executable does not get that treatment. Windows may still warn you the first time with a blue "Windows protected your PC" box because the app is not code-signed yet — click More info, then Run anyway. You only do that once.

• ALREADY ON WINDOWS? If the app currently does not start for you at all, updating cannot fix it — the updater is a .bat and is blocked by the same rule. Download the full build from the release page and unpack it fresh; your projects live in Documents and are untouched. If you update in place, an old app folder is left behind in your Cubric Vision folder — nothing uses it any more and it is safe to delete, about a gigabyte back. Updating from inside the app becomes the recommended way on Windows from this version onward — it no longer runs a .bat, so Smart App Control cannot block it. Getting TO 1.3.0 is a full download either way: the in-app updater in 1.2.0 and earlier is broken on Windows — it downloads the patch and then silently applies nothing — which is why this release ships no Windows update bundle. Windows patches resume at 1.4.0.

What is new

• NO GPU? SKIP THE INSTALL AND RUN IN THE CLOUD — setting up Cubric used to mean installing the ComfyUI engine first, even on a machine that could never use it. The install screen now offers a way straight past it, with a link to the setup video; Cubric then runs entirely on a cloud GPU. Turn it back off in Settings → RunPod whenever you want the local engine after all.

• NEW MODEL: FLUX.2 KLEIN — the fastest image model in Cubric Vision, and the only one that can take things out of a picture: paint over an object, hit Remove, and it is gone in about four seconds. It covers the full set in one model — generate from text, reshape an existing image, follow a depth reference, edit with up to three reference images, clean up detail and upscale — with eight style looks available on every one of those. Quality is modest next to Krea 2: reach for Klein when you want an answer now or something removed, and for Krea 2 when you want the finished piece. Tip: Depth on this model takes two images if you give it two — the first supplies the pose, the second supplies who is in it.

• KREA 2 EDITS FOLLOW YOUR INSTRUCTIONS MUCH MORE CLOSELY — the edit path has been rebuilt around the newer Krea 2 edit engine. Your reference image is now fitted to the shape of the picture you are making, so asking for a new setting actually gives you a new setting instead of quietly keeping pieces of the original background. Restaging someone into a new pose, outfit or scene lands closer to what you asked for, likeness is better, and references whose shape does not match your output no longer come out stretched or soft. Tip: the more of your reference frame the character fills, the better the likeness — crop your reference to the person before using it.

• KREA 2 QUALITY SETTING PRODUCES FAR MORE NATURAL IMAGES — everything made at the quality speed, not just edits. The sampler was pushing the model too hard, which flattened lighting, over-saturated colour and gave skin and surfaces a waxy, plastic look. It is now dialled back and a second detail pass finishes the image: lighting reads like a real photograph, skin and fabric hold real texture, and backgrounds resolve properly instead of going soft. It also sticks to your prompt more literally instead of quietly adding flattering details you never asked for. The fast setting is unchanged.

• BLEND TWO VERSIONS OF AN IMAGE WITH A MASK — when a re-run fixes one part of a picture and ruins another, you no longer have to choose. Paint a mask over the area you care about, select that entry and the one you want to borrow from, right-click and pick Mask composite. Add drops the other entry's pixels into the masked area; Subtract keeps the other entry and takes only the masked area from the one you painted. The blend lands as a new entry in the strip so nothing is overwritten, the edge is softened automatically, and a mask drawn as an outline fills in.

• HOLD ALT TO DRAG A GALLERY CARD STRAIGHT INTO ANOTHER APP — dropping a card into Discord, Photoshop or a browser upload box used to do nothing, because a plain drag only hands over a promise of a file that other apps ignore. Hold Alt while you drag and the card leaves as a real file, so any app that accepts a dropped file now takes it; you can let go of Alt once the drag is moving, and with several cards selected, Alt-dragging one drags them all. A plain drag is unchanged: onto the prompt to reuse a card, into a folder to save it.

• IMAGE DESCRIPTIONS ARE SHARPER AND GET STRAIGHT TO THE POINT — Describe an image no longer starts with "The image shows…"; the description opens on the subject itself so you can paste it straight into a prompt, and it is considerably more thorough, covering subject, clothing, materials, lighting, composition and background in enough detail to recreate the shot.

• PICK WHAT YOU'RE MAKING FROM A STRIP, NOT A DROPDOWN — every operation your model can do now sits above the prompt box as a row of chips with the current one lit. An operation you cannot run right now stays visible but dimmed, and hovering it says exactly what is missing — needs 1 image, paint a mask first, takes at most 2 images — instead of quietly disappearing from a list. The same strip appears inside the parameters popup, clearing the last image drops you back to text-to-image on its own, and the operation you actually chose comes back the moment you add an image again.

• EVERY OPERATION NOW EXPLAINS HOW TO PROMPT IT — operations want completely different prompts and nothing in the app said so. Open the parameters popup and there is now a ? above the operation strip: a short guide for whichever operation you are on — what the prompt is for, a couple of examples, and the mistake people usually make. It follows your model too: the SDXL family is pointed at comma-separated tags while the newer models are pointed at plain sentences.

• CHOOSING A MODEL IS A FULL-SCREEN CONTACT SHEET — the model dropdown is now a grid of preview tiles like the Model Library, so you pick by looking at the work rather than reading names — click the model button above the prompt and it opens the same way every time. LoRA & Upscale settings open from the model's own tile.

• TAB JUMPS BETWEEN THE GALLERY AND THE CARD YOU WERE WORKING ON — one key, both directions: from the gallery it opens the last card you had open in that project, and from inside a card it drops you back to the gallery. It remembers per project and after you close the app. If a project has no cards yet, Tab does nothing at all — and if you delete the card it was holding on to, it forgets rather than sending you somewhere you did not ask for.

• CROP OUTSIDE THE PICTURE — the crop box no longer stops at the edge of your image: drag it past any side and the new area fills with a colour you pick. That flat colour is what you hand to a model next — switch to an edit-capable model, do not mask anything, and just describe what should be there; it paints into the coloured area. Good for turning a portrait into a landscape. The box now also snaps to your image edges when it gets close, and there is a third Resolution mode beside Ratio and Free where you type an exact width and height and the result comes out at exactly those pixels. Images only; video crop is unchanged.

• MASKING IS REBUILT: FOUR TOOLS, AND IT CAN NOW FIND ANYTHING YOU CAN POINT AT OR NAME — Brush is painting, plain and simple. Points is new: click anywhere on the image and the mask snaps to whatever you clicked — a face, a pair of headphones, the coffee cup on the table; right-click to push the selection back off something it grabbed by mistake. Name lets you type what you want instead — hair, or cup and earring — say how many of each to find, and every one comes back as its own thumbnail to pick from. Detect is the face / hand / person finder you already had, unchanged. Everything adds on top of whatever you have painted, so all four mix freely. Anything a detector found now draws green, a new black-and-white view in the strip shows the mask exactly as it is handed to the model, the eraser's ring is visible against dark images, and the prompt box now stays open while you mask. Points and Name run on a new model that installs alongside the engine — expect a one-off download of about 1.7GB after you update. Tip: dots are cumulative, but they describe one region per run — click, hit Add, then start on the next part.

• CTRL+Z ON THE CANVAS — every mask edit can now be taken back — a brush stroke, a Clear, committing a detection — with Ctrl+Shift+Z to put it back again. The history follows you between mask tools, and starts fresh when you open a different image.

Fixes

• The app is no longer slow while connected to a cloud Pod. Opening your model library could leave the grid blank for ten to fifteen seconds, and Settings would sit there before showing how full your volume was or letting you press Disconnect. The cause was on the Pod, not in the app: measuring the volume's space and checking which models are installed were done in a way that made the Pod stop answer...

v1.2.0

Cubric Vision 1.2.0

Updating from 1.1.0? Your app checks for updates on startup, so it should offer this version in-app — one click closes it, fetches this release from GitHub, updates in place, and reopens. Your engine, models, and projects are kept. Prefer to do it by hand? Just run your install's update script — it pulls this version from GitHub for you.

What's new

• RunPod is dramatically faster — the two slow spots on a cloud engine are gone. Switching between image models no longer re-reads the weights from the slow network volume every time; they're staged to the Pod's fast local disk, cutting a cold model switch from about 2 minutes to a few seconds. And after a model downloads, the app no longer re-reads every weight back off the volume to checksum it — a completed multi-connection download is trusted directly, so the "Verifying…" step that could take longer than the download itself (about 3 minutes on a 2GB file) is now near-instant. Video models (LTX) that stream by design are unaffected.
• Stage all models on connect — we found that copying your installed models onto the Pod's fast local disk ahead of time makes the first generation start much sooner, so there's now a RunPod setting (off by default) that stages every installed model the moment the Pod connects. Off keeps the default, staging each model on first use.
• Depth Control for SDXL — the five SDXL generators (SDXL Realistic, SDXL NSFW, ILL Anime Beauty, ILL Anime, PONY Mix) can now follow the pose and composition of a reference image using a depth ControlNet. Pick the Depth operation, drop in an image, and the result keeps its structure while your prompt drives the content.

Fixes

• Dragged-out gallery cards keep their real name — dragging a gallery card out into a folder now saves it with its real filename from the project's Media folder (for example t2i_001.png) instead of a generic project-file.png, which also collided into an "already exists" prompt when you dragged out more than one. Dragging cards onto the prompt to reuse them is unchanged.
• Copy mask keeps the auto-detected regions — Copy mask only carried your brushed and erased strokes, so the auto-detected selection was lost on paste, and a mask made purely from auto-detect copied as nothing. It now carries the auto-detected regions too, matching what Download mask already exported.
• Completion toast fires reliably — a single generation finishing while the app is focused now shows the "Generation finished." toast; previously only multi-generation batches did. Notifications are quieter overall: routine confirmations (importing a model, uninstalling, adding cards to a project, switching models to continue a preview) no longer play a sound, and returning to the app after a download finished no longer replays a duplicate toast.
• Toasts moved to the bottom-right — the toast stack now stacks up from the bottom-right corner instead of the bottom-left, so notifications no longer cover your prompt text.

Engine

• ComfyUI updated to 0.28.0 (from 0.27), on both the local engine and cloud RunPod Pods. Brings a text-model sampling speedup and int8/int4 optimizations that mainly help local Turing / 16-series GPUs. No workflow or behaviour changes — every shipped model was re-swept and passed.

Platform status

• Windows — tested on the maintainer's Windows development machine. Not yet validated on a separate clean Windows host.
• Linux — install and launch validation only, on a weak Ubuntu laptop. Generation is unvalidated until a stronger Linux host or a contributor tests it.
• macOS (Apple Silicon / arm64) — artifacts are produced mechanically but are maintainer-untested; community validation is welcome. First launch of an un-notarized build: in Terminal run xattr -dr com.apple.quarantine "<folder>", then double-click start.command.

Cubric Vision is a local image and video generation app.