Wan2.2 Image-To-Video
---
🔗 Workflow Online Test Link
🔗https://www.runninghub.ai/post/1956812813727621122/?inviteCode=rh-v1171
First Click: Click the link to claim 1100 RH Coins (for new users only)
Uses of the Coins:
1. Can use RTX 4090 for free to render workflows for 2 hours
2. Allows generating approximately 20 videos (resolution: 1280*720)
---
📦 WAN Models to Download
---
🔴 Main WAN Model
Wan2_2-I2V-A14B-HIGH_fp8_e4m3fn_scaled_KJ.safetensors
🔗 [Download Link](https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/resolve/main/I2V/Wan2_2-I2V-A14B-HIGH_fp8_e4m3fn_scaled_KJ.safetensors?download=true)
🗂️ Place in: ComfyUI/models/diffusion_models
AND
Wan2_2-I2V-A14B-LOW_fp8_e4m3fn_scaled_KJ.safetensors
🔗 [Download Link](https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/resolve/main/I2V/Wan2_2-I2V-A14B-LOW_fp8_e4m3fn_scaled_KJ.safetensors?download=true)
🗂️ Place in: ComfyUI/models/diffusion_models
---
🟣 WAN2.2-LIGHTING
Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1-high_noise_model.safetensors
🔗 [Download Link](https://huggingface.co/lightx2v/Wan2.2-Lightning/resolve/main/Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1/high_noise_model.safetensors?download=true)
AND
Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1-low_noise_model.safetensors
🔗 [Download Link](https://huggingface.co/lightx2v/Wan2.2-Lightning/resolve/main/Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1/low_noise_model.safetensors?download=true)
🗂️ Place in: ComfyUI/models/vae
---
🟣 WAN VAE
Wan2_1_VAE_bf16
🔗 [Download Link](https://huggingface.co/Kijai/WanVideo_comfy/resolve/main/Wan2_1_VAE_bf16.safetensors?download=true)
OR
wan2_1_vae_fp8
🔗 [Download Link](https://huggingface.co/calcuis/wan-gguf/resolve/2a7520ea1e79d3f1f3f454a938613235633e1cba/wan2_1_vae_fp8.safetensors)
🗂️ Place in: ComfyUI/models/vae
---
🟣 WAN Text Encoder
umt5-xxl-enc-bf16.safetensors
🔗 [Download Link](https://huggingface.co/Kijai/WanVideo_comfy/resolve/main/umt5-xxl-enc-bf16.safetensors)
OR
umt5-xxl-enc-fp8.safetensors
🔗 [Download Link](https://huggingface.co/Kijai/WanVideo_comfy/resolve/main/umt5-xxl-enc-fp8_e4m3fn.safetensors)
🗂️ Place in: ComfyUI/models/text_encoders
---
⚠️⚠️⚠️ Prompt Optimization
This feature can only be used on online platforms. For local use, you need to disable the prompt optimization node.
---
⚠️ Torch Compile Warning
If your setup doesn’t support torch compile, set attention mode to sdpa in the model loader and bypas the Torch Compile settings and adjust the base_precision to just fp16
---
⚠️ Block Swapping
Block swapping helps if you have lower VRAM and/or get OOM. You can bypass it first and then update the blocks to swap to a higher number till you don't get OOM (up to 40) It will run slower so don't use if you have enough VRAM.
---
⚠️ Other known issues
If you get an error that has "FlowMatch" in it, please change your scheduler to uni_pc from FlowMatch_Causvid (or something else you like, dmp++_sde/beta is good too)
Description
FAQ
Comments (4)
its missing the "RH_LLMAPI_NODE" i cant find it anywhere ?
The "RH" node is exclusive to the online platform. When running locally, you can either disable this node and manually enter positive prompts, or replace the "RH" node by using nodes like "LLM Party" to call the API.
If you want to quickly test the effect, you can first run it on the online platform. The online platform offers 2 hours of free usage, which helps you quickly test the results. Online address: 🔗https://www.runninghub.ai/post/1956812813727621122/?inviteCode=rh-v1171
getting crisp videos with good prompt adherence using this workflow. and I like that it doesn't have too many 3rd party nodes - just a couple, as far as I recall.
8 second videos on RTX 4090:
848x1088, 8 total steps - as fast as 7 minutes.
720x960, 6 total steps - 3.8 minutes.
quality is great! I usually have to generate videos at lower resolution first, to "warm up" the model, otherwise getting OOM error right away.
any chance to adapt this workflow for use with some better LoRA loaders?
Yes! I am planning to train some LoRA models for use across different models. My plan includes conducting LoRA style training on Flux Kontext, Qwen Image, Wan 2.2, and more. Stay tuned!