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 the mask already on the image, so everything the mask tools can select — brush, detect, points, text, shapes, adjust — can drive a blend. Press Apply to keep it as a new entry at full resolution; your originals are untouched. The old blend asked you to decide before you could see anything, which meant running it three or four times to get one result.
Paint — draw on your image and let the model take it from there. A new Paint group with a brush, an eraser and a colour picker. A rough scribble tells the model where, how big and what colour — something a prompt alone cannot say. Paint a shape, then mask it and run Detail over it, and the model works from your sketch instead of guessing. It gets the same ten brushes as the mask — feather, airbrush, chisel, calligraphy, spray, charcoal, stipple, dry brush and the two rounds — and the same rectangle, triangle and ellipse, committed here as Fill or Erase in the colour you picked. Adjust works on paint too, and it is how you outline something: grow a scribble and the new edge is filled in your colour while what you already painted is left alone, shrink it and the edge is eaten inward with every surviving brushstroke keeping its own colour, or reduce it to a band and the scribble becomes its own outline. The paint stays on your image while you switch to the mask tools, and it keeps itself per image until you close the app, so nothing is lost while you work. Undo works on it like anything else on the canvas. Press Apply to bake it in as a new entry — your original is untouched, so deleting the new entry undoes it.
Resize — one click back to the original size. An Original Size button under the width and height boxes puts both back to the picture's own dimensions, so you can try a size, change your mind and start again without looking the numbers up.
Hover sound now has a volume, not just an on/off. A volume slider sits next to the gallery size slider, so you set how loud a hovered video or audio card plays without leaving the gallery. Sliding it to zero is the mute — which is why the "Play audio on hover" switch is gone from Settings; it was the same control, two rooms away. At zero an audio card stays silent instead of pretending to play, and a video still previews without sound.
MiniMax H3 — video that comes with its own sound. A new video model that generates the picture and synchronized stereo audio in a single pass, instead of handing you a silent clip to score afterwards. The sound belongs to the scene because it was made with it, not added over the top. It does Text to Video and Image to Video, and like LTX 2.3 it takes a first frame, a last frame, or both. Resolutions run up to 4K, though the sizes above its native canvas are final-render territory rather than something to iterate at: twice the pixels costs a little over three times the time, so the top of that ladder is for a card that can carry it, not somewhere to work. It is a big model — 50GB of weights, a slow generator, and a 12GB graphics card at the minimum — and it asks you to accept its licence before it downloads, which is covered further down. If what you want is the same character, place or voice across several clips, reach for its sibling below instead.
MiniMax H3 Reference — the same character, place or voice in every clip, with no training. A new video model that works from references instead of a first frame. Give it a face, a character sheet, a location, a clip whose camera move you want, or a voice, and it generates a new video that keeps them — up to nine images, three videos and three audio clips at once. Nothing you give it appears in the output as-is: these are references, not frames, so the model is free to put your character somewhere new rather than animating away from a picture. Point at a specific one in the prompt by typing @, which lists what you have staged and drops in its tag — "the woman from <Picture 1> walking through <Picture 2>". Each thumbnail wears the tag it became, so there is nothing to memorise. Reference detail decides how much of each reference it reads: Match is the fast default, Max keeps more and is the one to use for a character sheet, where it costs time on every step. It outputs video with synchronized audio like the rest of H3, and it installs alongside the existing MiniMax H3 sharing its text encoder and both VAEs, so it is one extra model file rather than a second full model.
Video previews are much better. While MiniMax H3 or LTX 2.3 works, the card now shows the actual clip in real colour, playing from the first frame as it refines, instead of the coloured blocks video models used to give you.
Engine updates now apply in place — seconds, not a fresh download. Updating the generation engine used to delete the whole thing and download it again from scratch: around eleven gigabytes of engine, every add-on and every Python package, even when the actual change was tiny. It now updates just the parts that moved, so a typical update finishes in seconds and costs you almost no bandwidth. The full reinstall is still there and still runs by itself whenever the update genuinely needs one — an add-on that was retired, a change to the engine's own Python, or anything that does not go cleanly — so you never end up on a half-updated engine.

New operations

control (v1.0) — copy the structure of a reference image. On Krea 2 and Chroma.
ref2v_ms (v1.0) — MiniMax H3 Reference: the same character, place or voice across clips, with no training.

Fixes

The LoRA and upscale pickers in Model Settings open again. Clicking a LoRA slot or the upscale model did nothing — the arrow flipped as if the list were opening, but no list ever appeared, so neither a LoRA nor a different upscale model could be picked. Affected Model Settings since 1.3.0.
Installing a model no longer fails after an app update changes a Python package. Once an update moved one of the engine's Python packages, the next model install died with Download Failed right after the files had downloaded — and every install after that failed the same way, so retrying could never work. The engine was running, and Windows will not let anything overwrite a file a running program is using. The update is now applied while the engine is stopped, on its next start.
Most model downloads now have a second route when your network blocks the first. Every model file was served from one address, so an ISP filter, a school or office network, or antivirus web protection sitting on that one address stopped the whole catalogue — nothing would install, and the app looked broken when the network was the problem. If the connection to the usual host fails, the download now retries the same file from Hugging Face automatically, and anything already downloaded is kept rather than started over. It also no longer loses the explanation: if both routes fail you still get the "your network blocked this, try a VPN" message instead of a bare error code. This helps when a network blocks the address; a network that interferes with the transfer itself can still break a download. A handful of files still have a single route: the three small preview decoders, which exist nowhere else in the form the app needs, and a few of the largest model files, which we serve straight from the host that publishes them.
Ctrl+scroll no longer changes the size of the interface. macOS sends a two-finger trackpad pinch as a Ctrl+scroll, and the app was treating that as the shortcut for changing UI size — so on a Mac, trying to zoom into an image blew up the entire interface instead. That shortcut is gone on every platform; use Ctrl (or Cmd) with + / -, which always did the same job. And the size you pick now survives a restart — it used to be forgotten on every launch, which meant anyone who needs a larger interface set it again every single time they opened the app.
The NVIDIA PiD upscaler now upscales, and no longer crops. A required setting had gone missing from its recipe, so the engine skipped the step that does the work and handed your picture back at its original size. The setting is back, the same gap has been swept for across every model, and PiD no longer crops the edges off an image it had to scale.
Updating from inside the app reopens it again. Pressing Update closed the app and left it closed, even when the update had applied perfectly — which read as a crash. The updater now reopens the app when it finishes, writes a log next to the app (update/update.log), and if an update fails it tells you why on the next launch instead of vanishing silently. Because the updater that runs is the one already installed, this takes effect when updating FROM this version — updating TO it still needs you to reopen the app yourself.
The startup splash now covers the whole wait, instead of vanishing on the slow starts it exists for. It was being closed the moment the window had something to paint — including the browser's own "cannot connect" page, which is what shows while the app's engine is still starting on a cold or slow first run. So the splash disappeared early, or on Linux failed to appear at all, and you watched an error page until the app arrived. It now stays up until the app itself is loaded.
A clear message when another program is using Cubric Vision's port. The app serves itself over a local port, and if something else already held it, Vision either opened onto whatever that other program was serving or sat on a black window with nothing to click — no menu, no refresh, no way out. It now stops and tells you which port is taken.
Running your own ComfyUI no longer breaks the app's. If you already had ComfyUI open, the app quietly used your install instead of its own, because both were on the same port. Yours does not have the app's custom nodes, so every generation stopped with "Node 'Input_Seed' not found. The custom node may not be installed." — which read as a broken install, when nothing was wrong with it. The app's engine now runs on its own port and stays out of the way of yours, so you no longer need to close ComfyUI before opening Cubric Vision. If something else is using that port, the app now tells you instead of silently generating on it.
Text detect finds things again. Naming an object found nothing at all whenever the result count was 1 — the default, and the setting most runs use. The count was written into the prompt in a form the detector read as part of the object's name, so it went looking for something that was never there. Raising the count to 2 or naming two objects had been dodging it by accident. Affected Text detect since 1.3.0.
Masked edits no longer fail on large images. The mask was being sent at a smaller size than the picture it belonged to, and any source over roughly 1500 pixels on its long edge stopped with a dimensions error. Affected FLUX.2 Klein's masked edit since 1.3.0.
Notifications no longer sit on top of what you're working on. They now appear at the top centre of the window, the one strip the interface keeps clear. Every corner they lived in before — bottom right, bottom left, top right — covered controls or your image.
The progress bar no longer keeps moving after you press Stop. Stopping a job that reports no percentage — an upscale, a detection — left the bar sweeping away under an IDLE label until the next job happened to reset it.
Typing a size in Resize no longer throws an error. Entering a width or height ran the preview again on every single keypress, against the half-finished number you were still typing — so the "1" of "1024" was treated as a one-pixel-wide image and stopped with a failure dialog. The preview now waits until you finish and click away.
The mask brush no longer skips on a fast stroke. The brush placed a dab at each position the mouse reported and nothing joined them up, so a quick drag left a dotted trail of gaps instead of a solid stroke — worst on a large image zoomed out, where the same hand movement covers far more of the picture between readings. Strokes are now continuous at any speed.
The live preview now looks like your picture, not a colour blob. While a generation runs, the thumbnail is decoded by a tiny preview model, one per model family — and the ones for FLUX.2 Klein were never installed, so it fell back to a rough colour projection that told you almost nothing about what you were getting. On macOS and Linux none of them were installed, so every model had the blob. They now install with the engine. A few models still show the rough preview on purpose: Krea 2, Qwen and Wan 2.2 share a preview decoder with a known bug that corrupts the real generation, so the mediocre preview is the safe choice until that is fixed upstream.
The batch count no longer appears where it did nothing. On SDXL's Image to Image and Control, asking for several images only ever returned one — the setting reached a part of the workflow those operations do not use. The control is now hidden there rather than lying about it. Text to Image is unaffected.
The home screen no longer claims you have no models when it cannot tell. On a cloud-only setup — the local engine install skipped, no Pod connected yet — the models count on the home screen read 0 / 18 on a machine whose drive was full of weights. Models live in a folder the local engine owns, so with no engine there is nowhere to look and nothing to count: the app was showing a zero it had never measured. It now shows a dash until an engine, local or cloud, can actually answer.
RunPod's "Stage all models on connect" is locked until you save an API key. The panel said the RunPod controls were locked until a key was saved, then let you flip that one switch anyway — a setting about connecting to a Pod you had no way to connect to yet.
The install screen no longer flickers during the dependency step. Two things report progress at once — the phase the installer is on, and the files it is fetching — and they were taking turns overwriting the same line, several times a second. Each now has its own place to speak.
Reuse Prompt now brings back settings you never touched. Reusing an older generation restored the controls you had changed, but silently skipped any control you had left alone — so a setting you had since nudged upwards stayed where it was instead of dropping back to what that run actually used. The app was only recording the settings you edited, and read a missing one as "leave it as it is". Every control a generation used is now recorded, whether you touched it or not. Older generations cannot be repaired — those values were never written down — but everything from this version on reuses exactly.
Models you never installed no longer show a half-finished download bar. Six model tiles were showing progress bars — 17%, 22%, 33%, 35%, 36%, 49% — for downloads nobody ever started, for two different reasons. Two of them were counting files that belong to a model you do have installed and billing them to its sibling; the other four were counting leftovers from models you had removed earlier, which nothing was clearing up. Both are fixed, and a bar now counts only the files that model would actually have to fetch.
A model that is part-way onto your disk no longer looks like it is downloading right now. A model whose files are partly present showed the same filling progress bar as a live download, so a library with a few of those read as several models installing at once — which never happens, since installs run one at a time. It now reads plainly as 98% ON DISK, beside the Install and Installed labels. A moving bar in the Model Library means one thing only: something is downloading this second.
Uninstalling now tells you the truth about what it removed. On a cloud Pod, the confirmation counted every file it asked about as deleted — including files that were never on the volume in the first place. One uninstall reported eight files removed when exactly one existed. And in the opposite case, a model whose files had already gone said "model files kept on disk; still installed" when nothing was kept and nothing was left. Both are gone: a file is only reported as removed when it was actually there and actually deleted.
A generation that fails before it starts now tells you why. Some failures happened before anything reached the engine, and said nothing at all: no message, no error, and the card simply sat on QUEUED while the queue counter climbed. Worse, that failed generation kept its slot, so every generation after it queued behind one that had already given up — the app looked frozen, and only restarting it cleared the queue. Those failures now release the slot and say what went wrong.
A failed model install now tells you why. When an install failed after the files had downloaded, the error box came up completely empty — no reason, nothing to act on, and no hint whether it was worth retrying. The reason was there in the logs the whole time; it just never reached the dialog. It does now.
A hiccup mid-download no longer throws away the download. A 25GB model that stalled for a minute — a router blip, a bad moment on the line — was declared failed outright, even though the gigabytes already on disk were perfectly good. It now retries on its own and picks up exactly where it left off. It still stops promptly and tells you what to do when your network is genuinely blocking the connection, rather than sitting there retrying something that will never work.
Uninstalling now clears up after itself properly. Uninstalling a model has always kept files that another installed model still needs, which is right. But once that other model went away too, its files were left behind for good — owned by nothing, deleted by nothing, and invisible in the app because a model you don't have installed offers no Uninstall button. On this machine that had quietly accumulated ~16GB. Uninstall now also clears out weights that no installed model needs any more. Files shared with something still installed are kept, exactly as before. This works on a cloud Pod's storage volume too — where it matters more, because that volume outlives the Pod and you pay for its size every month, so anything stranded on it was costing you until you noticed it.
The negative prompt now does something on LTX. Anything you typed into the negative box on an LTX video was being ignored. The usual mechanism for it only applies at a guidance strength LTX does not run at, and the one part that can carry a negative at LTX's settings had been pointed at your positive prompt by mistake — so it was nudging the video towards what you asked for instead of away from what you didn't want. It is wired correctly now, and leaving the box empty switches it off cleanly instead of leaving it half-applied.
Resize Video works on a cloud GPU whatever you have installed. It stopped with a missing-node error unless that cloud volume happened to have Wan 2.2 on it — the video node it needs only ever arrived as a side effect of installing that model, so a volume built up from any other model never got it. The nodes every tool relies on are now put in place when you connect, the way they already were on your own machine.
The Qwen Image Edit tier stays put. Quality/Turbo/Hyper was being remembered per operation, so switching between Edit and Control silently changed the tier underneath you. It is now one setting for the model.
Setting up the engine downloads less, and stops hiding a step that always failed. The generation engine is built from a set of add-ons, and each one used to install its own list of supporting libraries — over a dozen separate passes that re-checked the same libraries again and again, sometimes installed a package only to replace it moments later, and on some machines pulled in graphics-card libraries a computer without that card can never use. The app now installs one list, decided and tested in advance, in a single pass: less to download, less to unpack, and the same libraries on every machine instead of whichever add-on happened to go last. One of those add-ons also had a piece that failed to install on every computer, on every platform, on every fresh setup — and reported success anyway. It was never used by anything in the app, so it is gone.
The same extra model folder can no longer be added twice. On Windows, one folder can have two spellings — a long one and a short PROGRA~1-style one — and the app treated them as two different folders, so the same LoRA or upscale folder could end up in your list twice. Both spellings now resolve to the same folder.
Updating the engine no longer leaves tools quietly missing. After an engine update, five add-on packs could come back without the Python packages they need — the masking detectors, the pose and depth guides, several samplers and part of LTX. Nothing warned you. The tools simply were not there, and the first sign was a generation failing on a feature that had worked the day before. The engine now reinstalls those packages whenever it replaces itself, and a check was added so a future update cannot skip them the same way.
Installing to a cloud Pod that is still waking up no longer looks like a crash. Start a model install in the first seconds after a cloud engine comes up and every file could fail at once, behind a Download Failed box asking you to report it on GitHub — for something that fixes itself. The engine's storage service takes a few seconds longer to answer than the engine does, and the app was reading that gap as a real error. It now tells you the engine is not ready yet and to try the install again in a moment. Genuine failures — a bad file, for instance — still report exactly as before.
A cloud generation no longer disappears when a model adds new tools. Some models bring their own add-on packs, and the cloud engine has to restart to pick them up. It did that immediately — including while something else of yours was still generating. Whatever was running was killed mid-render and simply vanished: no error, no message, nothing in the queue, and the picture you were waiting for never arrived. Queue a batch, start one of those models, and the work already in flight was gone. The engine now waits for the work in progress to finish before it restarts, and if it cannot, it says so and leaves your generation alone instead of throwing it away.
A cloud install that was interrupted can be started again. If your cloud engine went away in the middle of a model install — stopped, deleted, or restarted — pressing Install again did nothing at all: no error, no progress, just a bar sitting where it stopped, and the only way out was restarting the app. The app was still treating that abandoned download as if it were running. It now checks with the engine first and restarts anything the engine is no longer working on.
Your cloud drive is no longer reported as full when it is not. If you had a download that was interrupted part-way, the app counted that half-finished file at its full final size — so a 21GB model that had only fetched 2GB was counted as all 21. The more unfinished downloads you had, the further out the number drifted, and the space bar in Settings read high for the same reason. Then the next install was refused for lack of room that was actually there: in one measured case the app said 39GB free and stopped an install while about 91GB was genuinely free. The message you got made it worse, because it told you to free up space — which was not the problem and would not have fixed it. The app now measures what the files really occupy on the drive rather than what they will occupy once finished.

Engine

COMFYUI ENGINE UPDATED TO 0.31.0 - from 0.29.2, two upstream releases in one step. It is a small update to install: the engine itself moves, but nothing in the heavy machine-learning runtime does, so there is no multi-gigabyte re-download.
MiniMax H3 generates faster on the new engine, and so do LTX 2.3 and Wan 2.2 - the work upstream did on H3 sped up the other two as well. Everything else runs as it did.
Every shipped model was re-swept after the update by executing a real generation for each of its operations on a GPU, 35 in total, and all of them passed.

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 2011 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.

Navigation Menu

Repository navigation

v1.4.0

Choose a tag to compare

Sorry, something went wrong.