Hugging Face's logo

Datasets:
Cseti
/
ComfyUI-Workflows

Dataset card Files
xet
 Community
ComfyUI-Workflows / wan /2.2 /upscaling /README.md
Cseti's picture
Cseti
Add Support section to all workflow READMEs
3a6cc54
preview code
|
raw
history blame contribute delete
4.22 kB

WAN 2.2 Upscaling v1

Model: WAN 2.2 A14B LowNoise (GGUF Q4_K_M) Type: Upscaling (Video-to-Video)

Per-subject crop upscaling pipeline for WAN 2.2 video output. Subjects and faces are detected automatically via SAM2 segmentation and Florence2 face detection, then cropped, upscaled with pixel-based models, refined with WAN video-to-video, and composited back. RIFE VFI is used for frame interpolation on the final output.

Preview

Preview v2 (looping)

Requirements

  • ComfyUI: recent stable build
  • Model: wan2.2/Wan2.2-T2V-A14B-LowNoise-Q4_K_M.gguf — download (place in ComfyUI/models/unet/)
  • VAE: wan/Wan2_1_VAE_bf16.safetensors — download (place in ComfyUI/models/VAE/)
  • Text encoder: umt5-xxl-enc-bf16.safetensors — download (place in ComfyUI/models/text_encoders/)
  • LoRA (speed): Wan/lightx2v/lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank64_bf16.safetensors — download
  • LoRA (quality, example): Wan/wan2.2_fun-reward/Wan2.2-Fun-A14B-InP-low-noise-HPS2.1.safetensors — reward LoRA for quality improvement; replace with any WAN 14B compatible LoRA or remove
  • SAM2 model: sam2.1_hiera_base_plus.safetensors — auto-downloaded by DownloadAndLoadSAM2Model
  • RIFE model: rife49.pth — auto-downloaded by RIFE VFI
  • Upscale models (place in ComfyUI/models/upscale_models/):
    • 4x-ClearRealityV1.pth — download — fast photo-realistic upscaler
    • phhofm/1xDeJPG_realplksr_otf.pth — download — JPEG artifact remover / enhancer
    • 1xSkinContrast-HighAlternative-SuperUltraCompact.pth — download — skin detail enhancer
    • DF2K_JPEG.pth — download — slower but higher quality upscaler
  • Custom nodes:

Notes

  • Face index: After running Florence2, each detected face gets a numbered ID visible above the bounding box in the preview. Set the Index field in the face crop node to the ID of the face you want to upscale.
  • crop_size_mult: Controls the crop area around the face. Use a higher value for faces that are large on screen, lower for small faces. Follow the preview output to calibrate.
  • VRAM: For 1080p output on under 16 GB VRAM, reduce context_frames to 33 or 41 in WanVideoContextOptions.
  • WAN denoise strength: Upscale passes use 0.3 denoise — low enough to preserve structure, high enough for texture refinement.
  • Input resolution: 832x480, 81 frames. The pipeline outputs at a higher resolution after upscaling.

Changelog

  • 2025-10-13 — Initial upload

Support

Producing and sharing this kind of open-source work requires renting cloud GPUs, which gets expensive quickly. If you find it useful and would like me to keep contributing, your support is very much appreciated:

Ko-fi Liberapay