1. Workflow Overview

This is a Wan model-based video depth control workflow specialized for video-to-video conversion. Key features:

• Depth map extraction from video frames

• Text-guided video stylization

• Two-stage sampling pipeline

• Automatic multilingual prompt translation

Core Models:

• Wan 2.1 T2V 1.3B: Video-optimized base model

• DepthAnythingV2: Depth preprocessor

• Florence-2-base: For auto captioning

• Wan Control LoRA: Depth adapter

2. Node Breakdown

Critical Components:

1. VHS_LoadVideo

   • Function: Load input video and extract frames

   • Requires: comfyui-videohelpersuite

   • Params: 16fps, 480x720 resolution

2. AIO_Preprocessor

   • Function: Depth extraction using DepthAnythingV2

   • Install: comfyui_controlnet_aux extension

   • Output: 512x512 normalized depth map

3. SamplerCustom (Dual-stage)

   • Process: 10-step high sigma + 15-step low sigma

   • Uses: Euler sampler

Special Dependencies:

• wan_2.1_vae.safetensors: From Wan model hub

• umt5_xxl_fp8: Multilingual text encoder

3. Workflow Structure

Group Logic:

• Video Input Group:

  • Nodes: VHS_LoadVideo → ImageResizeKJ

  • Function: Frame loading & normalization

• Depth Processing:

  • Nodes: AIO_Preprocessor → ImageScale

  • Output: Standardized depth maps

• Generation Control:

  • Contains: UNETLoader + LoRA loader + TeaCache

  • Key: 0.8 strength depth LoRA

• Two-Stage Sampling:

  • SplitSigmas → Dual SamplerCustom

4. Inputs & Outputs

Parameters:

• Required: Input video (e.g. "自动写提示词2.mp4")

• Optional: Positive prompts (auto-translated)

• Advanced: Depth control strength (0.08)

Output:

• MP4 video (16fps, H.264)

• Frame previews

• Translated prompts

5. Notes

• Hardware: Minimum 12GB VRAM

• Must install: VideoHelperSuite + ControlNet-Aux

• Model paths: All Wan models in wan/ subfolder

• Common issue: Frame rate mismatch causes audio sync problems

• Tuning: Lower CRF (current 19) for better quality