mickmumpitz/VACE_Skyreels_V3_R2V_Merge-GGUF

VACE Skyreels V3 R2V Merge — GGUF

GGUF quantized versions of Inner-Reflections/VACE_Skyreels_V3_R2V_Merge for use with ComfyUI-GGUF.

This is a merge of Wan 2.1 VACE with SkyReels V3 R2V by Inner-Reflections, enabling VACE-controlled reference-to-video generation. See the original model card for full details on capabilities and workflows.

Quantizations Available

Filename	Quant Type	Description
wan-14B_vace_skyreels_v3_R2V_e4m3fn_v1-Q4_K_M.gguf	Q4_K_M	Smallest, good balance of quality and size
wan-14B_vace_skyreels_v3_R2V_e4m3fn_v1-Q5_K_M.gguf	Q5_K_M	Recommended, better quality with moderate size

Usage

1. Install ComfyUI-GGUF custom nodes
2. Place the .gguf file in ComfyUI/models/unet/
3. Use the Unet Loader (GGUF) node in your workflow
4. For workflows, see the original VACE Skyreels repo and the VACE Phantom repo (same workflows apply)

Conversion Notes

• Quantized from the fp8 (e4m3fn) source weights
• This is a VACE variant with two 5D tensors (patch_embedding.weight and vace_patch_embedding.weight), which required a modified convert.py to handle correctly
• Conversion done using city96/ComfyUI-GGUF tools with a patched handle_nd_tensor method for multi-5D-tensor support

Credits

• Original model merge by Inner-Reflections
• SkyReels V3 by Skywork
• Wan 2.1 VACE by Wan-AI
• GGUF conversion tools by city96