Models, builds and LoRAs

What a model is, why one model comes in several builds, and why the only LoRA you ever handle is the one that holds a character's likeness.

Three words do most of the work in AI Models, and they are easy to run together. A model is the thing you choose. A build is which file of it you download. A LoRA is a small add-on file -- and almost all of them are already inside the workflows you run, so there is only one kind you will ever handle yourself.

A model is a name

A model is one card, with one name: Wan I2V 2.1, Flux Klein 2.0 4B, Qwen Image Edit 2511. The name is built to carry everything you would compare models on -- the line it belongs to, which release it is, what it makes, and how big it is -- so typing any part of it finds it. That is why the page has a search box and no filters on a model's make-up: there is nothing a filter could offer that the name does not already say. (The only toggles are the two that depend on your own setup -- Local and Cloud.)

Two things with different weights are two models. Wan 2.1 and Wan 2.2 are two models; so are Qwen Image and Qwen Image Edit; so are Flux Klein 2.0 4B and Flux Klein 2.0 9B. They may share a picture and most of a name, and they still are not interchangeable.

Size is worth reading twice, because it is the part of a name most often skipped. Two models can be identical but for the size, and nothing made for one of them works on the other.

A build is which file you download

One model is usually published as several files that trade memory for quality -- full size, a lighter fp8 build, a quantized GGUF one. Those are builds of one model, listed in the model's details, not models of their own. Flux Dev 1.0 is one card with an fp16 build and an fp8 build.

A build is known by its name -- what the makers call its precision and packaging -- and by the VRAM it wants. Both are in the Builds table, beside the download size and whether you already have that file.

A build is a hardware choice, never a look choice. Every build of a model makes the same kind of picture; the lighter ones fit a smaller graphics card and give up a little detail. See Model versions and quantization for how to pick one.

A large model is sometimes split into more than one file that are always loaded together -- Wan 2.2's 14B models come as a high-noise and a low-noise half, and both are part of the same build. You never choose between them; a build simply has two files instead of one.

Almost every LoRA is already in the workflow

A LoRA is a small add-on file that changes what a model produces. There are several kinds, and knowing which is which saves a lot of confusion -- but only one of them is something you handle:

Kind of LoRA Where it lives
Style -- Studio Ghibli, claymation, VHS, 90s animation Built into a workflow, by us. You get the look by picking that workflow.
Motion -- an orbit, a dolly, an explosion Built into a workflow, by us.
Speed ("turbo", Lightning -- the ones that let a model render in a few steps) Not used in OA's own workflows, which run at full quality. If a workflow you imported or edited does use one, every file it saves is tagged turbo.
Capability -- lip-sync, pose following, relighting Built into the workflow that needs it. Without it, the workflow would not do its job.
Identity -- one specific person, place or thing Yours. The only one you upload or choose.

So there is no list of looks to add to a model, and no shopping for one. A look is a workflow. When you want a different look, you pick a different workflow for the operation -- which is what a visual style recommends, once, for the whole look.

Identity LoRAs

An identity LoRA is trained on one character, one location or one prop, so that thing comes out looking like itself in every shot. It is made for exactly one model, and it only works with that model -- one trained for Flux Dev will not load on Wan.

Identity LoRAs live in the subject's own look gallery: from the Media gallery choose Upload identity LoRA..., say which model it was trained for and what its trigger word is, and mark it In use. OA then applies it to any run whose workflow has a free identity slot. Nothing about them lives in AI Models. See Characters.

Comfy words you may meet elsewhere

The wider ComfyUI ecosystem uses a lot of words OA does not. You do not need them, but if you read a tutorial or a model card you will run into them, so here is what each one becomes here:

What you will read elsewhere What it is here
Family (Wan, Flux, LTX) The first word of a model's name.
Version (2.1, 2.2) Part of a model's name. A new version is a new model.
Task (t2i, i2v, edit) Part of a model's name. What you ask for is an operation, which is not the same thing.
Size (5B, 14B) Part of a model's name.
Checkpoint, diffusion model, UNet The model's weight file -- one of the files a workflow needs.
Precision, quantization (fp16, fp8, GGUF Q4) Part of a build's name.
VAE, text encoder, CLIP vision, upscaler More files a workflow needs. OA installs them with it.
Negative prompt Built into the workflow.
Trigger word Part of the look a workflow carries, or part of an identity LoRA.
Custom nodes Something a workflow needs. OA reports a missing one like a missing file.

Learn more