ComfyUI error: Wrong VAE for Wan 2.2 5B

Load wan2.2_vae.safetensors in VAELoader. The Wan 2.1 VAE is only for the 14B Wan 2.2 models (per ComfyUI_examples / docs.comfy.org).

ComfyUI error: CUDA out of memory

ComfyUI docs state the 5B model should fit well on 8GB with native offloading; reduce Wan22ImageToVideoLatent length/resolution if still OOM.

Wan 2.2 text/image-to-video (5B)

Hybrid text-to-video and image-to-video with Wan 2.2 TI2V-5B (high-compression Wan2.2-VAE). Official ComfyUI template; consumer-GPU friendly with native offloading.

Estimated VRAM

~8 GB @ 720p — ✅ verified official figure. ComfyUI official docs: 'The Wan2.2 5B version should fit well on 8GB vram with the ComfyUI native offloading.' Wan-AI reference script separately states at least 24GB VRAM for 1280x704 TI2V with --offload_model True --convert_model_dtype --t5_cpu (not ComfyUI). Source, checked 2026-10-14.

Models required

wan2.2_ti2v_5B_fp16.safetensors — 10 GB ✅ verified 2026-10-14 (listing) · sha256 456f901338bd9eadbded3828b819109a9b68e8a525ca5cf8d0049a69fcfeca1e → models/diffusion_models
umt5_xxl_fp8_e4m3fn_scaled.safetensors — 6.736 GB ✅ verified 2026-10-14 (listing) · sha256 c3355d30191f1f066b26d93fba017ae9809dce6c627dda5f6a66eaa651204f68 → models/text_encoders
wan2.2_vae.safetensors — 1.409 GB ✅ verified 2026-10-14 (listing) · sha256 e40321bd36b9709991dae2530eb4ac303dd168276980d3e9bc4b6e2b75fed156 → models/vae

Tested GPUs

No verified test reports yet.

Nodes

UNETLoader CLIPLoader VAELoader Wan22ImageToVideoLatent CLIPTextEncode KSampler VAEDecode SaveAnimatedWEBP

Common errors & fixes

Wrong VAE for Wan 2.2 5B: Load wan2.2_vae.safetensors in VAELoader. The Wan 2.1 VAE is only for the 14B Wan 2.2 models (per ComfyUI_examples / docs.comfy.org). ✅ per source
CUDA out of memory: ComfyUI docs state the 5B model should fit well on 8GB with native offloading; reduce Wan22ImageToVideoLatent length/resolution if still OOM. ✅ per source

Low-VRAM tips

⚠️ General guidance, not individually verified.

Use ComfyUI native offloading (official docs: fits well on 8GB)
Keep resolution at template defaults for low VRAM
Fewer frames via Wan22ImageToVideoLatent length

Source & credit

Workflow by comfyanonymous (ComfyUI_examples) / Wan team. Get it from the original: https://comfyanonymous.github.io/ComfyUI_examples/wan22/. License: Wan 2.2: Apache-2.0.