Model families and version names

Model names look like they describe a tidy sequence: Wan 2.1 then Wan 2.2, Qwen Image Edit 2509 then 2511. They do not. A name usually packs together three separate things, and version numbers mean different things in different families. This page explains how to read a model name, so you can tell when two models are alternatives and when they are completely different tools.

A model name has three levels

Family -- who made it and what lineage it belongs to. Wan, Flux, Qwen Image, LTX Video, HunyuanVideo, SDXL. The family is the most useful part of the name, because add-on styles (LoRAs) and workflows are usually built for one family and will not work with another.

Version -- which release within that family. Wan 2.1 and Wan 2.2 are the same lineage at different points in time.

Build -- which file you actually download. One version is commonly published in several files that trade memory for quality, such as fp16, fp8, or GGUF. This is only relevant when you run locally, and it is covered in Model versions and quantization.

So "Wan I2V 2.2, fp8" is family Wan, job image-to-video, version 2.2, build fp8.

A bigger number is not always a newer model

Two numbering styles are in common use, and they look alike.

Sequential versions count up from the first release: SDXL 1.0, Flux 1 then Flux 2, Wan 2.1 then 2.2, HunyuanVideo 1.0 then 1.5. Here a higher number does mean newer.

Date-coded versions encode the year and month of release as four digits. Qwen's image models work this way: 2509 is September 2025, 2511 is November 2025, 2512 is December 2025. A higher number still means later -- but only within the same lineage, and this is where it gets genuinely confusing.

Qwen is the worked example, because its numbers invite exactly the wrong conclusion:

Model	Lineage	What it does
Qwen Image 2509	Qwen Image	Makes an image from a prompt
Qwen Image 2512	Qwen Image	Makes an image from a prompt (newer)
Qwen Image Edit 2509	Qwen Image Edit	Changes an image you already have
Qwen Image Edit 2511	Qwen Image Edit	Changes an image you already have (newer)

Qwen Image Edit 2511 and Qwen Image 2512 sit next to each other numerically, so 2512 looks like the newer version of 2511. It is not. They are different lineages doing different jobs, and 2512 is not an upgrade path from 2511 -- it cannot edit your image at all. The only meaningful comparison is 2509 against 2511 within the Edit lineage, or 2509 against 2512 within the image lineage.

The lesson generalises: compare version numbers only within one lineage, and check what the model is for before assuming one replaces another.

The same family often contains different jobs

A family name is not a promise that the models are interchangeable. Families frequently branch into siblings that do quite different work:

• Wan includes text-to-video and image-to-video, plus specialised siblings: Animate (drive a character from a performance), Fun Control (follow a control signal), Fun Inpaint (change part of a video), S2V (drive from sound).
• Flux includes the general image models plus Fill (fill in a region), Canny and Depth (follow an edge or depth map), Redux (re-use a reference image), and Kontext (instructed editing).
• Qwen Image splits into the image line and the Edit line described above.

Swapping one sibling for another is not a version change; it is a different tool. If a recipe asks for Wan Fun Control, Wan I2V will not do the same job.

Dev and distilled

Some families publish two flavours of the same version. Dev is the full model. Distilled is a version trained to produce a result in far fewer steps -- much faster, usually at some cost in detail. LTX Video publishes both for most versions.

Distilled builds suit draft and animatic work where speed matters more than finish; dev builds suit the final pass. Both are the same family and version, so add-ons for one generally apply to the other.

Base models and LoRAs

A LoRA is a small add-on file that adjusts a base model -- a style, a character, a lighting look. A LoRA is trained against a specific base, so it inherits that base's constraints: an SDXL LoRA needs an SDXL base, and a LoRA trained for one version of a family may not work on the next.

This is a real consideration when a family updates. Community add-ons are published per version -- you will see LoRAs named for the exact base they were trained on -- so moving to a newer version can mean leaving some add-ons behind until they are retrained.

Some newer models fold popular add-ons directly into the base model, so effects that previously needed a separate LoRA work with no extra file. Qwen Image Edit 2511 is one example.

Some models have no file at all

Several models in the catalogue run only as a paid service through an interface: Kling, Google Veo, Seedance, Gemini, the Flux Pro models, and others. There is nothing to download and your graphics card is irrelevant; you need an account and you pay per render. They appear alongside downloadable models because from your side the choice is the same kind of choice -- see Comfy Cloud or local?.

How to check rather than guess

• Each model's card links to the model's own page, which is the authority on what it does.
• Files needed, in the ComfyUI section of Settings, shows exactly which file a turned-on model requires, and how large it is -- the surest way to see that two similarly named models are in fact different files.
• Help me choose opens the assistant, which reads your project and hardware and recommends a specific model rather than a family.

Where these notes come from

Model lineages change, and anything comparative goes stale. The version-numbering behaviour described here was checked in July 2026 against the model publishers' own release notes -- the Qwen Image Edit 2511 release notes describe it as "an enhanced version over Qwen-Image-Edit-2509", which is what establishes that the Edit line and the image line are separate. Treat specific claims about which model is better as a starting point rather than a rule, and check the model's own page for the current picture.