rzgar/wan2.2_i2v_14B_bf16_hq

🤗 On Hugging Faceimage-to-videoapache-2.073 GBother✓ Checksum-verifiedupdated 0d ago
Magnet

WAN 2.2 14B I2V Hybrid Precision

  • FP32 Layers: All Attention mechanisms, Time/Condition Embeddings, LayerNorms, and the First/Last Transformer blocks remain in exact FP32. (preserves prompt adherence, color stability, and fluid motion)
  • BF16 Layers: The FFN/MLP are cast to BFloat16. (robust to BF16)

_Left: This model without LightX2V | Middle: This model with LightX2V | Right: BF16 | Fixed seed, Same LightX2V and prompt_

Your browser does not support the video tag.


Source: Wan2.2-I2V-A14B-FP32-ComfyUI.

LightX2V 4-Step LoRAs (BF16)

_Something is off with these LoRAs. I redownloaded the full FP32 models by Lightx2v, and the issues are still there: broken animations, body deformities, missing fingers or extra hands. Out of 10 videos, 2 are affected,so if you see it, you’ll know what it is._

Sampler: Euler with a Normal or Simple scheduler. (distill loras)

LightX2V 4-stepLoRAs, converted to BF16 (preventing accidental downcasting the base to FP16)