1. Workflow Overview

Purpose:
Automated batch processing for manga/novel promotion videos:

1. Video Frame Extraction → AI Tagging → Batch Repaint → Video Assembly

2. Supports SD1.5/SDXL & Flux models with multilingual handling

3. Counter-controlled batches for long videos

Core Models:

• Stable Diffusion 1.5/XL: Base image gen (SDXL-Anime | Sky_v3.1)

• WD14Tagger: Auto-tag frames (filters Chinese watermarks)

• ControlNet Tile: Preserves original composition (xinsir_controlnet_tile_sdxl_1.0)

• Stable Video Diffusion (SVD): Image-to-video extension

2. Critical Nodes

Model Dependencies:

• xinsir_controlnet_tile_sdxl_1.0

• WD14 Tagger

3. Workflow Structure

graph TB
  subgraph Input Group
    A[VHS_LoadVideo] --> B[ImageFromBatch]
  end
  subgraph Process Group
    B --> C[WD14Tagger]
    C --> D[CLIPTextEncode]
    D --> E[KSampler+ControlNet]
  end
  subgraph Output Group
    E --> F[VAEDecode]
    F --> G[SaveImage/VHS_VideoCombine]
  end

Group Breakdown:

1. SD1.5/XL Flow (Left):

   • Input: Video → Frames → Tagging → Repaint

   • Key params: 1024×768 res, Tile ControlNet strength 0.9

2. Flux Flow (Right):

   • Same logic with FP8 precision support

3. Image-to-Video (Top):

   • Optional SVD video generation

4. Inputs & Outputs

Input Parameters:

• Video: MP4, recommended res ≥512×256

• Counter settings:

  • count=0 (reset)

  • total=15 (frames per batch)

• Model: SDXL (default) or Flux

Output:

• Batch images: PNG in ComfyUI/output

• Video: GIF/MP4 (if VHS_VideoCombine enabled)

5. Important Notes

⚠️ Hardware:

• SDXL requires ≥12GB VRAM, Flux needs ≥8GB

• For long videos, use counter to split batches

⚠️ Troubleshooting:

• Chinese watermarks in tags: Check WD14Tagger's excluded_tags

• Blurry output: Adjust ControlNet's ending_control_step (default 0.9)

💡 Tips:

• Use BNK_CLIPTextEncodeAdvanced for better prompts

• Enable Lora Loader Stack for style LoRAs (e.g., anime)