Your browser does not support the video tag.

input image: Wan2_1_VAE_fp32 | VAE: Wan21_QwenHDR_hybrid_head_blend50

Hybrid VAE Collection

Experimental Wan 2.1 video VAE variants for use with Wan 2.2 or Qwen-Image.
Built by merging the native Wan decoder with weights from the Qwen Image HDR VAE.

All files share the same architecture as Wan2_1_VAE (16-channel latent, 4× temporal / 8× spatial compression). Drop-in replacements for the VAEDecode node — no workflow changes needed beyond picking a file.

What’s in this repo

Decoder fusions that keep Wan’s temporal decoding (correct video motion) while borrowing Qwen HDR detail and tone from the output head/decoder body. Use these when you want richer contrast, analog-ish character, or a touch of HDR punch without the slow-motion artifacts of the full Qwen HDR VAE.

Strength varies by file (head-only, partial blends).

Grayscale variants: the final RGB output layer is collapsed to luminance so every frame decodes as true black & white with Wan-correct motion.

Notes

• These are decode-time stylization VAEs, not a retrained model. Latent space matches the official Wan 2.1 VAE.
• Full Qwen Image HDR VAE alone can cause slow-motion / smeared motion on video, these hybrids avoid that by preserving Wan’s decoder body.
• Selective color ( red dress stays red, rest goes grayscale) is not possible at the VAE level, use post-decode color grading or a selective-color node after VAEDecode.

Credits

• Wan 2.1 VAE - Alibaba Wan Team / Comfy-Org
• Qwen Image HDR VAE - Felldude