WAN2.2_I2V_Notes

Use workflow

WAN 2.2 I2V GGUF COMPACT + SPEED WF | Lightning Lora 4+4 steps

Recommended model

Recommended resources

Start by generating the video at 720×944 or 512×880.
First, test prompt stability and whether the action LoRA loads correctly.
At 512×880, generation takes about 3 minutes on a RTX 4070 Ti Super.
Once prompts work and each model understands correctly, increase the resolution for final generation.
At 720×1024, generation time is about 15 minutes.
After producing a good base video, proceed with eye stabilization, upscaling, and frame interpolation.
To further improve stability, resolution, and smoothness, use the WanFaceDetailer workflow to do it all at once:
https://www.reddit.com/r/StableDiffusion/comments/1n5zwzu/wanfacedetailer/
After any upscaling model, remember to run Color Match (KJ nodes) for color calibration.
After color matching, use JPEG Compression Removal - FBCNN for artifact removal.
See FBCNN’s principles here:
https://github.com/jiaxi-jiang/FBCNN
Finally, perform frame interpolation with RIFE VFI (recommend rife47 or rife49).
Use either the rife47 or rife49 model for interpolation.
Do a final pass to check details, then publish.

Notes:

wan2.2_i2v is very sensitive to input image quality; higher input resolution tends to yield better results.
At 512×880, eyes often blur; at 720×944, eye stability is acceptable.
Using 720×1024 makes details much clearer.

使用工作流

WAN 2.2 I2V GGUF COMPACT + SPEED WF | Lightning Lora 4+4 steps

WanFaceDetailer

推薦模型

Wan2.2 I2V A14B GGUF

推薦資源

Wan 2.2 Lightning LoRAs

先用720944或512880的解析度生成影片

先測試提詞穩定度跟動作LORA載入是否正常

以這個512*880解析度生成，以RTX 4070ti super生成大約3分鐘左右

測試確認提詞正常且倔任模型可以正常理解後再提高解析度做生成

解析度提高到720*1024大約生成時間是15分鐘