AI Knowledge Central
patreon
WAN 2.2 On Any GPU (4–32 GB): Exactly What To Download + Where It Goes
🕑 Added 2025-08-27 18:00:57 +0000 UTC🚨UPDATE 29 OCT 2025🚨 - The latest update of ComfyUI broke the cg-Everywhere nodes I am using in most of my newer workflows. I have got your back: https://www.patreon.com/posts/emergency-wan-2-142314304
Hey friends! 💛
I’ve put together a dead-simple guide so you can get Wan 2.2 (A14B, GGUF) running without guesswork. You’ll need two models (HighNoise + LowNoise), a Text Interpreter (UMT5-XXL), and a VAE. Grab everything here:
💻Software:
ComfyUI | Generate video, images, 3D, audio with AI
🏎️ Sage Attention Install Guide:
https://youtu.be/q4jhieMwNkU
Sage Attention One Click installer for my Patrons:🎁
https://www.patreon.com/posts/edit-31-08-2025-135625892
👉 Repo:
Text 2 Video
https://huggingface.co/QuantStack/Wan2.2-T2V-A14B-GGUF/tree/main
Image 2 Video
https://huggingface.co/QuantStack/Wan2.2-I2V-A14B-GGUF/tree/main
What to download (match to your GPU VRAM)
4 GB: 14B isn’t realistic. Use the 5B variant instead.
https://huggingface.co/QuantStack/Wan2.2-TI2V-5B-GGUF/tree/main
8 GB: Q3_K_S or Q3_K_M for both HighNoise + LowNoise. Keep the Text Interpreter on CPU.
12 GB: Q4_K_S (or Q4_1) for both. Text Interpreter Q5_K_M (CPU or GPU if you’ve got room). VAE can be the smaller one if you’re tight.
16 GB: Q5_K_S (or Q5_0 / Q5_1) for both. Text Interpreter Q5_K_M on GPU, Wan 2.2 VAE recommended.
24 GB: Q6_K for both. Interpreter on GPU, Wan 2.2 VAE.
32 GB: Q8_0 for both. Interpreter on GPU, Wan 2.2 VAE Or go for the Fp8 version from Kjai
Rule of thumb: pick the same quant for HighNoise + LowNoise. Start conservative (Q4/Q5) and step up if you’ve got headroom.
Download these add-ons
Text 2 Video Lroa (Lightning x2v, 4-step)
High-noise: wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise.safetensors
Low-noise: wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise.safetensors
Image 2 Video Lora (Lightning x2v, 4-step)
High-noise: wan2.2_I2v_lightx2v_4steps_lora_v1.1_high_noise.safetensors
Low-noise: wan2.2_I2v_lightx2v_4steps_lora_v1.1_low_noise.safetensors
VAE
Text Encoder (UMT5-XXL)
LoRa's used:
Golden by T2Video:
https://civitai.com/models/1671285/retro-90s-anime-golden-boy-style-lora-wan-22-14b
WAN2.2_Spatial Magic—空间魔法:
https://civitai.com/models/1867349/wan22spatial-magic
Comments
Chris Wenzl
There is LongCat in ComfyUI for Videos that should be able to do even longer videos. I haven't tested it yet. So I can't say if its good or not.
Shelby Black
How do I keep the videos from getting darker as we extend them?
Nick Klimek
I'm having a hard time. When I load your workflow, the 2 gguf loaders in the top left, when I click the play triangle button, it says undefined, now I know you said that means they aren't in the correct folder, but I downloaded both the high and low, example "Wan2.2-T2V-A14B-LowNoise-Q5_K_M.gguf" and placed them both in the diffusion models folder. I have an nvidia 4070 graphics card, if that makes a difference. Thanks for all you do! Also, do any GGUF files go in to the folder "unet"? On the github, I believe it instructs me to put them in the 'unet' folder, but in your video it seems he doesn't talk about any files that should be in "unet'. You say to put them in "diffusion models" right?
Chris Wenzl
Wan 2.2 does not need a clip vision model. The model doesn't have an image cross attention layer. If you provide clip embeds via the clip vision model in the workflow, they are simply ignored as there's nothing to process them in the model.
sachin paul
clip vision?
