AI Models
The model catalogue -- every model, what it is for, what to read about it, and whether its files are on your machine.
AI Models is the model catalogue, and it belongs to you rather than to a project. Open it from your account menu, at the foot of the sidebar. It answers three questions:
- What is there? Every model we know about, one card each, with what it can do.
- What should I read about it? The model's own pages -- knowledge-base articles, videos, blog posts, the upstream source -- listed in its details, one press away.
- Where does it stand on this machine? Available on Comfy Cloud, on your own disk, or neither; and whether it will run on your graphics card. Then: install what is missing.
Nothing on it is per project, which is why it is not a project page. It works with no project open at all.
There is nothing to turn on here either. No switches, no on/off. A project starts using a model when you pick it for an operation, which happens where the work does: in a gallery's Recipe panel, press Choose an operation and pick from the models listed for the job in front of you. See Choosing the right model. If the shot is in a visual style you can have that look recommend the model you picked, so it is preselected next time.
You can also reach this page from See every model at the bottom of the operation list. Arriving that way, the page gives you a back arrow to the gallery you came from.
Getting set up
The first time through, a short setup card guides the choices that come before picking models, one question at a time:
- Whether you have a Comfy Cloud account you want to use.
- Whether you want to run models locally -- already set up, want help setting up, or cloud only.
- If local: a prompt to connect your models folder (see below).
It only asks what it doesn't already know -- if you have a Comfy Cloud key saved, or your models folder is already connected, it skips straight past those questions and won't ask again. Once you are set up, the card steps aside. Not sure which models to pick? Help me choose models, at the left of the toolbar above the list, opens the assistant, which reads your project and hardware and recommends a model.
What is on a card
A card is deliberately short. It carries a thumbnail, its provider's logo, and three things:
- The model's name, on up to two lines.
- The operations it can do (Lipsync, Image to video, Upscale, and so on).
- Where it can run: a small Local | Cloud strip, with each side solid green when the model can run there -- on your own ComfyUI, or on Comfy Cloud. Point at it to read the same thing in words ("Local: installed. Cloud: available"). A paid hosted service shows only Cloud.
Those three are what you compare cards by. Everything else about a model -- its builds, what it needs, what to read, what the licence allows -- is in its details, which you get by pressing the card.
That operations line is the one to read when two cards look alike. Some models are published as more than one set of weights, one per job, and each set is its own card -- same name, same picture, different jobs. Picking the wrong one leaves the job you actually wanted with nothing behind it.
One card per model, not per file. A model published at several precisions -- full size, a lighter fp8 build, a quantized one -- is a single card, and its details list the builds. Which build a project runs for a given operation is recommended by a visual style, or chosen when you run it. See Model versions and quantization.
The strip is a fact, not a switch. Nothing on a card can be turned on or off; it reports where the model can run as things stand. When a name or the operations line is too long for the card, point at it to read it in full.
Cards are ordered by Recommended first, then by how many Comfy workflow templates use the model (most-used first), then by name. Whether you have a model installed does not change its place.
Model details
Press a card to open its details. (Tab to a card and press Enter to do the same thing.) One panel, everything about the model:
- What it is for -- each operation it covers and what that operation does, the look it produces, and whether it is a hosted service.
- Builds -- one row per build, with its name, its download size, the video memory it wants, where you stand with it (installed on your ComfyUI or not, whether it fits your graphics card, whether Comfy Cloud has it) and its own Install. This is the part a single card could never say: the lighter build may be on your disk and fit comfortably while the full-size one is neither -- and they are different downloads, so Install belongs to the row and not to the model.
- Art styles -- what the model can look like. See below.
- Learn about it -- the pages that model's makers and our own writers have published: knowledge-base articles, videos we rate, blog posts, the HuggingFace card, the Comfy.org workflow it came from, the upstream repository. Curated data rather than a fixed list, so they grow as we write and find more, and the same links appear on the model's row in Files needed -- "what is this thing" gets asked in both places. A model with nothing written about it yet simply has no such section.
- Licence -- what the catalogue records. Some models are published under a non-commercial licence, and the panel says so; where we have nothing recorded, it says that instead of implying the model is unrestricted.
- Used in this project -- whether the open project has settled on the model for an operation, with a link to the workflow it runs in its Workflow Library.
There is no single Install for the whole model: a model is a name, and a name is not something you can download. Each build row carries its own.
There are no hover tips on the cards. Everything they used to whisper is here, in full, and reachable by keyboard.
Art styles -- what a model can look like
A card used to show one picture, which told you nothing about the question you were actually asking: can this model do anime, photoreal, watercolour? The Art styles in a model's details answers it, with pictures.
Each art style is a look we built for that one model and rendered real examples of -- the words that model needs to draw it. (A visual style is different: it is yours, and says what a scene means. When you add a recipe to a visual style you choose an art style for each model the recipe uses.) An art style card carries:
- The art style's name -- the look's common name: Anime, Comic, Claymation, Watercolour. The model's own plain offer is the model's default art style.
- Its primary example. Tap the picture to see every example we rendered, each with the full prompt and the settings behind it.
- The builds its workflows come at -- Builds: fp8 -- so you can see straight away whether it is one your card can run.
- Its prompt trigger text: the words that carry the look.
We build model styles; you cannot add one. They come with the product, and they grow as we test more models. Every model with a workflow also has a Model default look -- that workflow with no style words added -- so the plain offer is always there to pick. (That is a look on the model's card, not a style of your project.) A model we have not written any looks for simply has no style library section.
A model style is not a visual style. A model style answers what can this model look like?; a visual style answers what does this project look like here? One model style can feed several visual styles, and one visual style can mix models.
Using a look
The card's menu, top-right, holds Copy fragment -- it takes the prompt trigger text so you can paste it into a prompt by hand. A model card no longer approves a look for a visual style: a visual style lists the recipes for each kind of media, each with an art style for every model, on the Visual styles page.
Some looks we have rendered have no workflow yet -- the examples are there, the fragment is there, but we do not ship a workflow for that model. Those cards say so.
Used in this project
When you have a project open, a tick and a coloured border mark the models one of that project's visual styles recommends, and the model's details say so -- for example Recommended by: Anime (image to video) -- with a link to the Visual styles page, where you edit it. It is a report, never a second editor -- and pressing the tick opens that project's Workflow Library with the workflow highlighted, so "used" is followed immediately by "used for what". The model's details panel says the same thing in words, and carries the same link.
With no project open there are no ticks, because there is no "here" for a model to be used in.
How the list is ordered, and what is not on it
One grid, ordered by the ones we recommend first, then the most widely used, then the rest alphabetically -- the same order the operation picker uses, so the two lists agree about what is worth looking at.
You narrow it by typing. The search box above the grid matches a model's name, and a name is built to carry the line it belongs to, which release it is, what it makes and how big it is -- so "flux", "i2v", "2.2" and "14B" all find what you would expect. There are no filter buttons for those, because a button for something already in the name is a second way to ask one question. See Models, builds and LoRAs.
Beside the box sit the only two controls on the toolbar, Local and Cloud. They are not facets of the catalogue but facts about you: turn one on and the grid keeps only the models you could actually run that way right now. Both off shows everything.
There are no LoRAs anywhere on this page. A style, a camera move or a fast-render trick is built into a workflow, not bolted onto a model, so the way to get a look is to pick the workflow that has it. The one LoRA you will handle yourself is the identity LoRA that holds a character's likeness, and it lives in that character's own look gallery. See Models, builds and LoRAs.
Nor are helper files listed. Some files a workflow loads -- decoders, text encoders, the upscalers that sharpen an LTX render -- are parts of how a model does its job rather than models you could pick instead. They are installed with the workflow that needs them and appear in Files needed, but they are not cards here.
Installing what a card needs
The button in a card's top-right corner -- and the Install on a build row in the details panel -- takes you to the ComfyUI section of Settings, scrolled to that build's rows in the Files needed list and highlighted: its file, its size, where it belongs, where to download it from, and where to read about it. That is the only place files are fetched, so there is one set of actions to learn rather than two. The button stays on the card so that coming here to install something does not mean reading about it first.
Install on a build row asks for THAT build. Press Install on Flux Dev's lighter build and Files needed is headed "Flux Dev 1.0 -- fp8" and lists the fp8 file: not the full-size one, which is a different download of a different size. The card's own button means the build the catalogue ships, which is the one almost every model has only one of.
That works for any model, used or not. The list shows the build you arrived asking about whether or not any project has ever run it, so pressing "install this" never lands you on an empty table.
Example: MiniMax H3 and MiniMax H3 Singularity
MiniMax H3 appears as three cards, one per set of weights: MiniMax H3 (image to video and first/last frame), MiniMax H3 Reference to Video (reference images, video and audio) and MiniMax H3 Singularity. The operations line tells them apart.
MiniMax H3 Singularity is a community fine-tune of the reference-to-video weights (Apache 2.0, by WarmBloodAban on HuggingFace), aimed at sharper faces at a distance, less oily skin and cinematic action. It is local only -- Comfy Cloud does not carry its files -- and it offers three builds:
| Build | Size | Notes |
|---|---|---|
| w4a8 | about 11.8 GB | The one to run on a 24-32 GB card. On an RTX 5090 a clip took about half the time of the official int8 build. |
| int8 pruned | 21 GB | The same size as the official reference-to-video weights. |
| int8 | about 34 GB | Unpruned. |
It shows up as a choice for References to Video, next to the official H3 workflow. Neither carries a Lightning LoRA, and the output size is yours to set: 1344 x 768 (the official 768p) by default, with a warning on the panel when the duration and size together would cross the point where H3 ghosts (about 95 frame-megapixels). Say in the prompt that the speaker is alone unless you want others, and that the background stays identical to the first picture -- the prompt, not the reference, decides the scene.
References to Video (MiniMax H3, Spin with Mask) is a further choice for References to Video: give it one picture of a posed character on a plain light grey background and it makes a clip in which the camera circles the character, with the colour frame on the left and its mask on the right in every frame of the one video.
H3 reference mode is for references to video, not for a talking shot. It follows the audio or the picture, never both: measured on a real shot, the storyboard did not hold in 17 of 17 controlled takes and every take came out darker than the picture. So Lipsync no longer offers an H3 workflow driven by a recorded clip. The only H3 reference workflows with an audio input take a voice sample (Voice Sample to Video), never the recorded line. A talking shot on H3 speaks from its first frame instead: Talking First Frame (an Image to Video workflow) takes the storyboard as the first frame, which stays exactly as drawn, and speaks the line from the prompt; then Voice Transform Shot gives that clip the character's own voice -- from the voice sample set on the Character Voice page, which a talking shot requires (the page can make one from the character's TTS voice, or take an upload). The clip-driven models below (Wan S2V, InfiniteTalk, LTX Lip Dub) stay available under Lipsync for a project that wants them.
Every OA workflow runs at full quality. A speed-up (turbo, Lightning) LoRA damages parts the model has to invent, so the slow run is the intended one. One workflow still uses a distilled sampling path: LTX 2.3 Lip Dub, because its makers publish it only for their distilled model.
Example: Wan S2V 2.2
Wan S2V 2.2 makes one character speak a line from a single picture and a recording of the line: the voice drives the mouth, the face and a little head and shoulder movement. It is a choice for Lipsync (Lipsync (Wan S2V 2.2)), next to the InfiniteTalk and LTX Lip Dub workflows.
- Best for a single speaker framed close, especially realistic faces. It is weak at lip-sync on anime and cartoon characters -- for those, let MiniMax H3 speak the line from the first frame (Talking First Frame, then Voice Transform Shot).
- Say the emotion in the prompt, and ask for it to be held for the whole clip ("keeps the same broad, energetic smile throughout"). Left to itself it drifts back to a neutral face; H3 holds the picture's expression better. The workflow's default prompt already asks for this -- change the expression to match your line.
- The clip is as long as the line, up to 14.5 seconds, and the mouth is still when the speech stops. The model renders in chunks of about 5 seconds, and OA only renders as many as the line needs: a line up to 4.9 seconds is one chunk, up to 9.8 seconds two, longer three. A short line therefore costs about a third of a long one in GPU time. The workflow then cuts the result to Output Duration.
- It is delivered at 24 frames per second. The model itself renders 16; the workflow smooths that up to 24 so it matches the rest of your timeline, which can smear very fast mouth movement slightly.
- Full quality only. The workflow runs 20 steps and no speed-up LoRA. A fast 4-step version is possible but loses motion and quality.
- Builds: fp8 scaled (15.27 GB, the one the workflow loads) and bf16 (30.35 GB, a little less degradation). The workflow also needs the RIFE frame-interpolation and Video Helper Suite custom nodes for the 24 fps step.
- One speaker at a time. There is no way to say who is talking, so a shot with two speakers needs one run per speaker, or a different model.
Example: Wan InfiniteTalk 2.1
Wan InfiniteTalk 2.1 is a choice for Lipsync (Lipsync (Wan InfiniteTalk 2.1)), beside Wan S2V 2.2. It is an audio-driven add-on that runs on top of the Wan I2V 2.1 weights, so it is listed as a model of its own rather than as Wan I2V 2.1.
- Runs on Comfy Cloud and on your own ComfyUI. The workflow uses only core ComfyUI nodes, and every file it loads is available on Comfy Cloud.
- The quality choice. Both InfiniteTalk and Wan S2V 2.2 lip-sync well. InfiniteTalk is sharper with more natural movement; S2V is softer with less movement.
- One speaker. The build OA runs is the single-speaker fp16 one. A multi-speaker variant exists but OA does not use it.
- The clip is as long as the line, up to 9 seconds, at 25 frames per second, and the mouth is still when the speech stops. A longer line is cut off.
- Full quality only. The workflow runs 40 steps and no speed-up LoRA, with the settings InfiniteTalk's authors publish for full quality. That makes it slower than Comfy's own 6-step template. The negative prompt is used: left blank, it is Wan's standard list of things to avoid (over-exposure, a static frame, blur, subtitles, badly drawn hands and faces).
- What to install (local). The InfiniteTalk patch, the Wan I2V 2.1 480p weights, the umt5 text encoder, the Wan 2.1 VAE and the wav2vec2 audio encoder. If one is missing, OA tells you which and where to put it.
Will it run on your machine?
Once your account knows your graphics card, a model's details answer that question for each of its builds -- while you are choosing, not after you have downloaded several gigabytes:
| What a build says | What it means |
|---|---|
| Runs on your card | It fits in your video memory and will run at normal speed. |
| Slow on your card | It only fits by spilling out of video memory, so expect it to be slow. |
| Too big for your card | Beyond what your card can manage. Comfy Cloud can still run it when it is available there. |
| Hosted service -- nothing to install | It runs on the provider's hardware, so your card does not come into it. |
The verdict is deliberately absent in one case: when your card is not known, because a fit we never checked is exactly the claim this is meant to prevent. Where the catalogue has no memory figure for a build, it says so rather than guessing -- and that never stops you running it. Your graphics card and video memory live in your account settings -- see Choosing the right model.
Files needed
As soon as a project you have open wants files you don't have locally, this page tells you so and offers to take you to the ComfyUI section of Settings, where getting files actually happens. That is where the Files needed list lives.
The list comes from the looks your project uses. A visual style recommends which workflow runs each operation, and a workflow names the exact files it loads -- including which build of a model. So the list is what those workflows need and nothing else: a project whose Lipsync look recommends the lighter LTX build is asked for that build's file, never for the full-size one. Each section is headed by the build it is for, and under it sit that build's own file plus the helper files its workflows load alongside it.
For each file the list shows:
- Its name and size, with a running total, so you can confirm you have the disk space before any multi-gigabyte download.
- Its status:
- Installed -- already there.
- Your file -- you have told the platform to use a file of your own for this.
- Found nearby -- you already have this file, just not where the platform expects it (see below).
- Not installed -- the platform can fetch it for you.
- Manual -- no automatic download; you have to source the file yourself and put it in the folder named.
The File details button on a row shows its full install path, where to download it from, and -- under Learn more -- the same reading list the card offers. The row's "..." menu is where the doing happens: Install (download now) when the platform can fetch it, Download it to open the download link yourself, and Choose a file on disk... to point at one of your own. You choose what to install; nothing large downloads without you asking for it.
Files you already have
If you already run ComfyUI, you probably have many of these files under your own names and folders. When the list finds one that isn't where the platform expects, it says Found nearby and offers Use the file found on disk on that row's menu instead of proposing a duplicate download. Your file is then referenced exactly where it sits, changing nothing on disk, so any of your own workflows that point at it keep working, and the row reads Your file. Use the expected file instead undoes that.
You can also point at a file the scan did not find with Choose a file on disk....
The same offer reaches you from a run. When a local run finds that one of the files its operation names is sitting in a sub-folder of your models folder rather than at the top of it, the report says Found nearby: <file> in <folder> -- use it, and pressing it points that run at your copy. It only ever matches the same filename: a sibling version of the same model is a different file, and swapping one in would quietly change your pictures. Like every other note about your folders, it belongs to your account and not to the project, so a teammate's layout is their own business -- see Choosing the right model. Everything you have answered this way is listed under Settings, in the ComfyUI section, as Files you pointed us at -- that is where you undo one after moving a file back; see Local ComfyUI Configuration.
Connecting your models folder
Connecting your ComfyUI models folder is a one-time setup that lets the platform see which files you already have and install missing ones. It lives under Settings, in the ComfyUI section, with a short walkthrough of how to find your ComfyUI folder. Once connected, it is remembered across visits, and the AI Models page uses it automatically. This works in Chromium browsers such as Chrome and Edge; elsewhere the Files needed list still shows you the links and folders to install by hand.
Finding a model
Type into the search box. A model's name carries everything you would look for -- the line it belongs to, the release, what it makes, how big it is -- so "wan", "2.2", "i2v" and "14B" all find Wan I2V 2.2, and "flux" brings back every Flux model at once. That is the whole reason there are no filter dropdowns: a facet could only offer what the name already says.
Two toggles sit beside the box, and both are facts about your machine rather than about the catalogue:
- Local -- show only models installed on your local ComfyUI
- Cloud -- show only models available on Comfy Cloud
They combine as AND with the text. Neither narrows the list to one project: this page is your catalogue, not a project's inventory. To see what a project has settled on, open its Workflow Library -- and a card the open project uses still says so, with a link straight to the workflow it runs.
Refreshing local availability
The circular-arrow button at the right of the toolbar ("Recheck local ComfyUI installation") re-probes your local ComfyUI install to update the Local indicators -- which model files and which API-node integrations are present -- so you get an accurate picture without leaving the page. It is disabled, with the reason, when local ComfyUI is not running.
Learn more
- Models, builds and LoRAs -- what each of those words means here
- Choosing the right model -- how to pick for your goal and hardware
- Workflows -- the project's library, and what it settled on per operation
- Model versions and quantization -- which version fits your card
- Comfy Cloud or local? -- the trade-off in detail
- The Media Gallery -- how a chosen model becomes an operation you can run