Wan 2.2 I2V A14B on 10 GB of VRAM
Download Q2_K (5.3 GB). With the model's working memory it needs about 9.6 GB, leaving 0.4 GB spare on 10 GB. Quality: heavy loss; a last resort. For better quality, Q3_K_M (7.2 GB) also runs, with about 1.5 GB spilling into system RAM — a little slower, still practical.
Worth trying on NVIDIA GPUs: with an up-to-date ComfyUI, Dynamic VRAM (on by default for NVIDIA since March 2026) streams whatever does not fit from system RAM, and ComfyUI's own start-up message recommends native FP8/INT8 files over GGUF, saying they “will be faster even if they are larger than your memory”. So before settling for a heavily compressed GGUF, try the native FP8 file (14.3 GB). The file recommended above is the best one that fits entirely — the safe choice on older ComfyUI versions, AMD and Intel. Comfy blog: Dynamic VRAM →
Using Kijai's WanVideoWrapper instead of the native nodes? Its WanVideo BlockSwap node keeps part of the model in system RAM. To run the FP8 file (14.3 GB) on 10 GB, start with blocks_to_swap ≈ 26 (of 40). Each block is about 0.4 GB; raise the number if you still run out of memory, lower it for speed. estimate WanVideoWrapper →
01What to download
| Part | File | Folder | Size | |
|---|---|---|---|---|
| Model | Wan2.2-I2V-A14B-HighNoise-Q2_K.gguf high-noise model (early steps) Q2_K · QuantStack/Wan2.2-I2V-A14B-GGUF | models/unet | 5.3 GB | Download → |
| Model | Wan2.2-I2V-A14B-LowNoise-Q2_K.gguf low-noise model (late steps) Q2_K · QuantStack/Wan2.2-I2V-A14B-GGUF | models/unet | 5.3 GB | Download → |
| Text encoder | umt5_xxl_fp8_e4m3fn_scaled.safetensors UMT5-XXL FP8 (scaled) · Comfy-Org/Wan_2.1_ComfyUI_repackaged | models/text_encoders | 6.7 GB | Download → |
| VAE | wan_2.1_vae.safetensors Wan 2.1 VAE · Comfy-Org/Wan_2.2_ComfyUI_Repackaged | models/vae | 0.3 GB | Download → |
| Total download · keep about the same free on disk | 17.6 GB | |||
Alternatives: UMT5-XXL FP16 (11.4 GB). The 16-bit text encoder is a little more faithful; the smaller one is chosen here because it loads faster and needs less RAM. Sizes read from Hugging Face (2026-09-25). “Download” links start the file directly; the file name links are the same files the official ComfyUI workflows use.
System RAM: 32 GB or more. ComfyUI keeps the model files, the text encoder and the VAE in system RAM and moves them to the GPU as needed. With this set of files that is about 17.6 GB, plus roughly 6 GB for Windows, ComfyUI and a browser: 23.6 GB in total. With less RAM it still runs, but Windows starts swapping to disk and loading gets very slow. calculated
02Every file, on 10 GB
| File | Size | Needed | On this card | Quality | Download |
|---|---|---|---|---|---|
| 16-bit SAFETENSORS · Comfy-Org | 28.6 GB | 32.9 GB | 22.9 GB too big | the original weights | Hugging Face → |
| Q8_0 GGUF · QuantStack | 15.4 GB | 19.7 GB | 9.7 GB too big | practically identical to the original | Hugging Face → |
| FP8 SAFETENSORS · Comfy-Org | 14.3 GB | 18.6 GB | 8.6 GB too big | practically identical to the original | Hugging Face → |
| Q6_K GGUF · QuantStack | 12.0 GB | 16.3 GB | 6.3 GB too big | very close to the original | Hugging Face → |
| Q5_K_M GGUF · QuantStack | 10.8 GB | 15.1 GB | 5.1 GB too big | close; small differences in fine detail | Hugging Face → |
| Q4_K_M GGUF · QuantStack | 9.7 GB | 14.0 GB | 4.0 GB too big | good; some loss in fine detail and text | Hugging Face → |
| Q3_K_M GGUF · QuantStack | 7.2 GB | 11.5 GB | spills 1.5 GB | noticeable loss of detail | Hugging Face → |
| Q2_K ← GGUF · QuantStack | 5.3 GB | 9.6 GB | fits · 0.4 GB spare | heavy loss; a last resort | Hugging Face → |
Sizes from the Hugging Face file listing, read 2026-09-25. “Needed” = file + 3.5 GB working memory for this model + 0.8 GB kept free for the system. Wan 2.2 A14B uses two files of this size (high-noise and low-noise); only one sits in VRAM at a time, both must fit in system RAM.
03The text encoder
UMT5-XXL + CLIP Vision H: 11.4 GB as 16-bit, 6.7 GB as FP8, 3.7 GB as GGUF Q4_K_M (plus CLIP Vision H (1.3 GB) for the input image). ComfyUI encodes the prompt first and can push the encoder out of VRAM before sampling, so it does not have to fit together with the model. On 10 GB the FP8 encoder fits on its own, so prompt encoding stays fast.