Image and video models are measured differently from language models. There is no context to budget for: what decides whether a model runs is the size of the files ComfyUI loads, stage by stage, and how much memory is left for the work. Every figure below comes from those files. Of the 15 models here, 10 load ComfyUI's own template set with room to spare on a 24 GB card, and 13 do with the smallest file that card can use.
Every model, compared
Sorted by what ComfyUI's official template needs at its largest stage. "Smallest card" is the smallest card listed on this site that loads the template set with room to spare (under 85% of its memory).
| Model | Type | Template peak | Download | Smallest card | On 24 GB | Licence |
|---|---|---|---|---|---|---|
| Anima | image | 4.4 GB | 5.6 GB | RTX 5070 | BF16 | non-commercial |
| FLUX 3 Image Klein · Klein (distilled) | image | 6.7 GB | 17.8 GB | RTX 5070 | BF16 | non-commercial |
| MOSS-Image · SFT | image | 7.3 GB | 13.0 GB | RTX 5070 | BF16 | Apache 2.0 |
| FLUX.2 · klein 4B | image | 8.0 GB | 12.5 GB | RTX 5070 | BF16 | Apache 2.0 |
| Fibo 2.0 · Dev | image | 9.4 GB | 27.0 GB | RTX 5070 | BF16 | non-commercial |
| Qwen-Image-2.1 · Text to image | image | 9.5 GB | 26.8 GB | RTX 5070 | BF16 | non-commercial |
| Z-Image · Turbo | image | 12.7 GB | 20.7 GB | RTX 5080 | BF16 | Apache 2.0 |
| Krea 2 · Turbo | image | 13.4 GB | 18.6 GB | RTX 5080 | FP8 scaled | Krea 2 Community License |
| Wan 2.2 · Text to video, A14B | video | 14.5 GB | 35.6 GB | RTX 4090 | FP8 scaled | Apache 2.0 |
| HunyuanVideo 1.5 | video | 19.2 GB | 29.0 GB | RTX 4090 | FP16 | territory limits |
| Qwen-Image-Edit-2511 | image edit | 20.8 GB | 30.2 GB | RTX 5090 | FP8 mixed | Apache 2.0 |
| LTX-2.5 · Distilled | video + audio | 24.3 GB | 44.9 GB | RTX 5090 | streams | LTX-2 Community License |
| MiniMax H3 | video + audio | 24.4 GB | 40.1 GB | RTX 5090 | W6A8, pruned | territory limits |
| FastH3 · Omni, references to video | video + audio | 25.6 GB | 41.3 GB | RTX 5090 | streams | territory limits |
| Pika H3 Flex · Text or frames to video with audio | video + audio | 26.4 GB | 42.0 GB | RTX 5090 | W6A8, pruned | territory limits |
Check any model on your own card, or pick a precision by hand, in the ComfyUI VRAM checker. Wan 2.5, 2.6, 2.7 and 3.0 are missing on purpose: none has open weights (checked 2026-10-02). No card big enough? Here is every way to run ComfyUI in the cloud, compared by what you pay.
The models
Anima
An anime image model from CircleStone Labs, made with Comfy Org and built on NVIDIA's Cosmos-Predict2 2B. Template peak 4.4 GB; the smallest file loads on an RTX 5070.
FLUX 3 Image Klein
Black Forest Labs' open FLUX 3 model from October 2026: a 5.6-billion-parameter image model in two versions, the step-distilled Klein that samples in 16 steps and serves Black Forest Labs' own API, and Klein Base, the undistilled one for fine-tuning and LoRA training. Template peak 6.7 GB; the smallest file loads on an RTX 5070.
MOSS-Image
OpenMOSS's open text-to-image model from October 2026: a 5-billion-parameter single-stream transformer with a Qwen3.5 9B text encoder and the FLUX.2 small VAE. Template peak 7.3 GB; the smallest file loads on an RTX 5070.
FLUX.2
Black Forest Labs' second FLUX generation: dev (November 2025), a 32-billion-parameter model, and klein (January 2026), 4B and 9B models built for consumer cards. Template peak 8.0 GB; the smallest file loads on an RTX 5070.
Fibo 2.0
The second generation of Bria AI's FIBO, from October 2026: an 8-billion-parameter model that generates and edits images with one set of weights, from a prompt and up to four reference images. Template peak 9.4 GB; the smallest file loads on an RTX 5070.
Qwen-Image-2.1
Qwen's image model from September 2026: a 7-billion-parameter generator with native 2K output, driven by an 8-billion-parameter Qwen3-VL text encoder. Template peak 9.5 GB; the smallest file loads on an RTX 5070.
Z-Image
Tongyi-MAI's 6-billion-parameter image model, one of the few here under Apache 2.0. Template peak 12.7 GB; the smallest file loads on an RTX 5070.
Krea 2
Krea's open image model from June 2026: a 12-billion-parameter transformer in two versions. Template peak 13.4 GB; the smallest file loads on an RTX 5070.
Wan 2.2
Alibaba's Wan 2.2, from July 2025, remains the newest Wan with open weights. Template peak 14.5 GB; the smallest file loads on an RTX 5080.
HunyuanVideo 1.5
Tencent's 8.3-billion-parameter video model from November 2025: 480p and 720p, 121 frames by default, with a separate super-resolution model that takes the result to 1080p. Template peak 19.2 GB; the smallest file loads on an RTX 5070.
Qwen-Image-Edit-2511
Qwen's image-editing model from December 2025: give it an image and an instruction. Template peak 20.8 GB; the smallest file loads on an RTX 4090.
LTX-2.5
Lightricks' open video model from August 2026, a 22-billion-parameter transformer that generates sound with the picture. Template peak 24.3 GB; the smallest file loads on an RTX 5090.
MiniMax H3
MiniMax's open video model from August 2026, which makes the soundtrack with the picture: stereo audio, 4 to 15 seconds at 24 frames a second. Template peak 24.4 GB; the smallest file loads on an RTX 4090.
FastH3
FastVideo's eight-step distillations of MiniMax H3, from Hao AI Lab at UC San Diego: V2 (September 2026) for text and first-and-last-frame video, and Omni (October 2026), which takes up to nine images, three videos and three audio clips as references. Template peak 25.6 GB; the smallest file loads on an RTX 5090.
Pika H3 Flex
Pika Labs' post-trained MiniMax H3 from October 2026: two LoRA adapters, one for text and first-and-last-frame video and one for up to 12 references, that run on the MiniMax H3 base and generate video with synchronised audio in six sampling steps. Template peak 26.4 GB; the smallest file loads on an RTX 4090.
How the numbers work
A ComfyUI workflow runs in stages: the text encoder turns your prompt into embeddings, an optional prompt enhancer rewrites it, then the diffusion model samples and the VAE decodes. Models load when a node needs them and move out when another needs the room, so what has to fit is the largest stage, not the sum of every file. Wan 2.2's A14B models are two experts that take turns, so one is on the card at a time and both are downloaded.
What is left after the weights is working memory, and that is the part nobody publishes per model: it grows with resolution, frame count and batch size, and it depends on the attention kernel. So each model page quotes what the vendor says it needs, word for word, beside the file arithmetic, and says which figure measures what.
- File sizes: exact byte counts from the Hugging Face API, for Comfy-Org's repackaged repos and the vendors' own ComfyUI-ready files, read on 3 November 2026.
- Template sets: the official workflow guides on docs.comfy.org, linked from each model page.
- Card memory: manufacturers' specifications, as on each GPU page. NVIDIA cards only; ComfyUI also runs on AMD and Apple silicon, which we have not sized.