Choosing the right model
How to pick an AI model for your goal, your budget, and your hardware -- and how the assistant helps you decide.
There are a lot of AI models, and "which one should I use?" is genuinely hard -- it depends on what you are making, where you want to run it, and what your computer can handle.
Decide at the point of use, not in advance
You do not have to choose your models up front, and you should not try. Which model suits your artwork is only really knowable by running one on it. So the choice lives where the work happens: in a gallery's Recipe panel, press Choose an operation, find the thing you want done, and the models that can do it are listed right there -- each with what it is good for, and what it will cost on your machine. Pick one and get on with it.
A visual style lists recipes, so you choose once. If the shot is in a visual style -- "Normal", say -- and that style lists recipes for what you are making, the first one that fits the shot is already selected. If the shot has no style, or its style lists nothing for it, every model for the operation is listed and you pick. A pick made in the gallery is for this run only; the style's recipes are set on the Visual styles page.
A scene can look different, without setting anything on its shots. Give Scene 4 a style called Dream whose rough-clip list starts with a recipe built on a different model, and every shot in Scene 4 starts with that recipe -- and no other scene changes. A style is optional: a scene or shot with none adds no look and leaves the choice to you. Styles are defined at the project, in Pick the recipes, and the assistant sets them on scenes and shots when it plans them. See Visual styles.
Nothing is turned on anywhere. There is no switch to find first. Every model the catalogue has an operation for is listed, and whether it can run right now is worked out for you and written on the row -- see "Which models you can run" below.
Two other pages sit either side of the choice, and neither of them is the chooser.
- The Visual styles page is where each look's recommended recipes live, in order, for each kind of media, with the art style of every model they use. Go there to see them, change one, or make another look.
- AI Models, in your account menu, is the model catalogue: every model there is, what to read about each one, and whether its files are on this machine. Nothing on it belongs to a project. Go there to learn and to install. It also has a Help me choose models button that opens the assistant, which asks about your project and hardware and recommends a specific model.
Everything below is for when you want to understand the field rather than just get a shot made.
Start from your goal
Pick the kind of thing you are making; that narrows the field far more than any spec sheet:
- Still images -- character art, backgrounds, concept frames. SDXL and Flux are common starting points; SDXL has a huge community behind it and a wide range of looks, while Flux tends toward sharper, more literal results.
- Editing an image -- changing part of a picture, restyling, or combining references. These are separate "edit" models rather than the ones you use to make an image from scratch.
- Video -- turning an image or a prompt into a short clip. This is the most demanding kind of model. Video models differ a lot in speed and final quality: some are fast but softer, others are slower but hold detail better. Because reports vary and change over time, treat any "X is better than Y" claim as a starting point, not a rule -- and lean on Comfy Cloud for video if your own card is modest.
- Upscaling -- making an existing image or frame larger and cleaner. These are small, specialised models.
- Voice and speech -- turning text into spoken dialogue, or cloning a voice. See the audio pages for how these fit the dialogue pipeline.
A model's NAME is built to answer most of this: it carries the line it belongs to, the release, what it makes and how big it is, so typing any part of it finds it. Read the whole name before comparing two candidates -- a bigger version number does not always mean a newer model, and two cards can differ only in the size at the end. Models, builds and LoRAs explains what each part of a name means.
Then decide where it runs
Every model can run one of two ways, and the choice changes what "which model" even means:
- Comfy Cloud -- the model runs on rented hardware. Nothing to install, and you always get the full-size model, but you pay for each render.
- Local ComfyUI -- the model runs on your own computer. Free to run once set up, and fully under your control, but you download multi-gigabyte files and your graphics card has to be big enough.
For the trade-off in detail, see Comfy Cloud or local?.
For local, your graphics card decides the version
Here is the part people miss: a single model is usually published in several versions that trade memory for quality and speed. A card with less memory can still run a big model -- you just use a smaller version of it. Which version fits your card is the whole game for local generation, and it is explained in Model versions and quantization.
You almost never have to choose the version by hand: a recipe names the workflow, and so the build, each of its steps runs, locally and on Comfy Cloud alike. Where the catalogue holds the same model packaged another way, a recipe on a style's list offers Change build, and the build you pick is kept with that recipe. It is never quietly substituted for a different one -- a smaller version of a model is a different picture, not the same picture made cheaper.
Which models you can run
Whether a model can run for you right now is arithmetic, not a setting, and it is worked out per row:
- Cloud is available when the model is in the Comfy Cloud catalogue and you have Cloud switched on.
- Local is available when every file the operation needs is already installed AND the build fits your graphics card.
"Already installed" means the exact files, by name. Having a version of a model is not the same as having the one this build opens: a smaller, lower-precision copy of Qwen Image Edit will not do for a build that wants the full-size file, and saying it would is how a render gets to the point of failing before anyone notices. So the row checks the filenames the operation actually asks for, and if one of them is missing it says so.
A build that does not fit your card is Cloud or nothing -- you are never quietly given a smaller, different model than the one you chose, and never asked to run something your card cannot really manage.
Unless the same model comes in a build that does fit. Then the row's menu offers Other builds, and choosing it opens the build choice on that one. It is the ordinary choice, for the run you are making. It appears only when both halves are true -- the build the row is offering will not run on your card, and another build of the same model will -- so it is never a question about builds you did not need to think about.
A model you cannot run is still listed, last, with the reason and the way to fix it on the row: Local: 3 files, 21 GB to install (which takes you straight to that model's files), or Cloud: switch Cloud on in Settings. Nothing is hidden from you, because knowing what exists is part of choosing.
Nothing about installing is pushed at you before then. The download list opens from the row you were looking at, at the moment you chose it -- not when you open the operation list, and not from Try, which is for finding out whether you want the model at all.
The first time you run one on your own machine
The first local run of an operation checks that the files it names are really there, before anything is submitted.
If your ComfyUI models folder is connected (see AI Models), the check reads the folder and, if something is missing, stops and tells you which file -- by name -- with a link to install it. That is the whole point: a missing file used to be discovered minutes later, by ComfyUI, in a message written for a developer.
Two things can fix it from that report:
- Install the missing files takes you to the Files needed list, with sizes and download links, focused on the model you were running.
- Found nearby:
<file>in<folder>-- use it appears when you already have that exact file somewhere else under your models folder. Press it and this run points at your copy where it sits. Nothing is moved, nothing is downloaded twice, and the operation itself is unchanged -- it is a note about your own folders, kept with your account, so it does not follow the project to anyone else's machine.
If your models folder is not connected, the run goes ahead: you are trusted about your own disk. If ComfyUI then refuses the job over a file, you are told which file it was, in words, with the same install link -- rather than having to read it out of a ComfyUI error.
Once a check comes back clean, that operation does not ask again on this machine.
Choosing a build
When you pick a model for an operation in a gallery, you may be asked which build to use. A model is often published in more than one form -- a full-size file, a smaller one at lower precision -- and they are not interchangeable: they want different amounts of memory and produce slightly different pictures.
The choice lists every build your project could run it with, described by what it actually is rather than by a filename: its precision or packaging, how much there is to download, how much graphics memory it needs, and its Runs on line for your own machine. Two ways out:
- Use -- runs that build, for this run. To make a style use it every time, set the style's recipe on the Visual styles page.
- Cancel -- decide later. To compare builds by looking, see "Deciding later, in an experiment" below.
Most models come in exactly one build, so the choice is usually a one-line statement of what it is rather than a list.
Where a model does come in two, both are offered as separate entries under the same operation and the same model -- one row, two builds -- and the row names the build it is offering ("bf16, 43 GB, needs 48 GB VRAM") so you can see which one you are looking at. The names tell them apart too, with the build after a semicolon: Lipsync (LTX Video 2.3 Dev; bf16) and Lipsync (LTX Video 2.3 Dev; fp8). Which one the row offers first is the model's usual build, not whichever was added most recently.
Deciding later, in an experiment
Choosing a model is sometimes a thing you can only do by looking. Start an experiment from the pictures you want to try it on, try the candidates there, and star the one you liked best. An experiment is a sandbox: nothing in it changes what a style runs. To make a style use the winner, start the experiment for a kind of media on the Visual styles page (Try the top 3) -- its Decide puts the winning recipe on that style's list -- or put a recipe built on the winner on the list yourself.
The assistant can help: in an experiment, asking which result won is enough. It never decides on its own -- a recommendation is not a choice, and a project's answer to "which model for this" should be one you can remember making.
When a model "does the wrong job"
Some models are published as more than one set of weights -- one per job -- and each set is a separate card with the same name and the same picture. If a model you expected is missing from an operation's list, check the operations listed on its card: the one you picked may simply do a different job.
How the platform helps
- Choose an operation, in any gallery's Recipe panel, is the chooser. A table of contents of the five goals comes first, then the operations under each, then the models for the job -- in one order, with the ones this project already uses first. Each row carries a Runs-on line worked out for your own graphics card: whether it fits, how slow it would be, how big the download is, and what a cloud run has actually cost you. See The Media Gallery.
- The Visual styles page is where each look's recommended recipes live, and where you define the looks themselves.
- AI Models, in the account menu, is the catalogue, not the chooser: every model, its reading list, and whether its files are on this machine. A card takes you to the ComfyUI section of Settings, where the Files needed list shows exactly which files the build you asked for wants, how big they are, and which you already have -- so you only ever download what you actually chose.
- Help me choose models, on the AI Models page, opens the assistant, which reads your project and your saved hardware and recommends a model that fits. It will not make the choice for you: when an operation has more than one model ready it asks which, and when it has none it offers to open the list or to try a few side by side in an experiment.
- The assistant is honest rather than salesy: it will tell you a model needs a big download, or will not fit your card, or costs money per render on the cloud.
Model quality, speed, and cost all move over time as new versions ship. Use these notes to get oriented, then check the model's own page for the current picture -- its File details in the Files needed list links to the model's HuggingFace card, the workflow it came from, and any tutorials.