🛠️ Workflow Overview

Purpose and Function:
This workflow utilizes AI to repair blurry videos by performing frame-level restoration and enhancement. It combines video loading, frame processing, video control embeddings, VAE encoding and decoding, sampling, Lora models, and video synthesis to achieve high-quality video restoration, re-rendering, or style migration.

Core Features:

1. Video Loading and Parsing: Extracts frames from the video and pre-processes them.

2. Video Enhancement and Re-rendering: Repairs blurry frames or applies new visual styles using AI models.

3. Video Synthesis and Output: Combines the processed frames into the final video output.

🔥 Core Models

1. Wan2_1-T2V-1_3B_bf16.safetensors

   • Function:

     • The core model used for video generation and repair. It enables video re-rendering or high-definition restoration.

   • Installation:

     • Place the .safetensors file into the models/WanVideo directory.

     • Alternatively, install it via ComfyUI Manager.

2. Video Lora Model: Video-to-Video Control

   • Model Name:

     • 视频转视频控制_wan2.1-1.3b-control-lora-tile-v0.2_comfy.safetensors

   • Function:

     • Enhances video restoration and style consistency during processing.

   • Installation:

     • Place the Lora model into the models/Lora folder.

3. VAE Model: Wan2_1_VAE_bf16.safetensors

   • Function:

     • Encodes and decodes video frames, generating high-resolution textures.

   • Installation:

     • Place the VAE model in the models/WanVAE directory.

⚙️ Nodes Explanation

1. VHS_LoadVideo

   • Function:

     • Loads and parses the video into frame images, audio, and metadata.

   • Parameters:

     • Frame loading cap: 50

     • Skip first frames: 0

     • Select every nth frame: 1

   • Output:

     • Frame images, audio, frame count, and video information.

2. WanVideoVAELoader

   • Function:

     • Loads the VAE model for encoding and decoding frames.

   • Parameters:

     • Model path: Wan2_1_VAE_bf16.safetensors

   • Output:

     • VAE model data.

3. WanVideoModelLoader

   • Function:

     • Loads the Wan video generation model.

   • Parameters:

     • Model path: Wan2_1-T2V-1_3B_bf16.safetensors

   • Output:

     • Video generation model.

4. WanVideoLoraSelect

   • Function:

     • Loads the Lora model to enhance video consistency.

   • Parameters:

     • Lora model path: 视频转视频控制_wan2.1-1.3b-control-lora-tile-v0.2_comfy.safetensors

   • Output:

     • Lora model data.

5. WanVideoTextEncode

• Function:

  • Encodes text prompts into embeddings.

• Parameters:

  • Positive prompt:

    a chinese woman dancing, dress

  • Negative prompt:

  • oversaturated tones, overexposure, static, blurry details, subtitles, grayish image, worst quality, low quality, JPEG artifacts, ugly, deformed limbs

• Output:

  • Text embedding data.

1. WanVideoSampler

   • Function:

     • Samples video frames using the embeddings and model data.

   • Parameters:

     • Sampling steps: 30

     • CFG scale: 5

     • Sampling method: unipc

   • Output:

     • Enhanced video frames.

2. WanVideoDecode

   • Function:

     • Decodes latent space images into video frames.

   • Parameters:

     • Resolution: 272x272

   • Output:

     • Restored or generated frames.

3. VHS_VideoCombine

   • Function:

     • Combines the video frames into a final video with optional audio.

   • Parameters:

     • Format: H264-MP4

     • Frame rate: 16

     • CRF: 19

   • Output:

     • The restored video file.

🧩 Workflow Structure

✅ Group 1: Video Loading and Parsing

• VHS_LoadVideo → Loads and parses the video into frames.

• WanVideoVAELoader → Loads the VAE model.

• WanVideoModelLoader → Loads the video generation model.

• WanVideoLoraSelect → Loads the Lora model.

✅ Group 2: Text and Image Embeddings

• WanVideoTextEncode → Encodes the text prompts.

• WanVideoControlEmbeds → Processes image embeddings.

✅ Group 3: Video Generation and Restoration

• WanVideoSampler → Samples frames based on the embeddings.

• WanVideoDecode → Decodes the sampled frames into video images.

✅ Group 4: Video Synthesis and Output

• VHS_VideoCombine → Merges the frames into the final video.

• Output format: MP4

🔥 Inputs & Outputs

✅ Inputs:

• Blurry video file.

• VAE and video generation models.

• Lora model.

• Positive and negative text prompts.

• Video sampling parameters.

✅ Outputs:

• High-definition, restored video.

• Option to include audio.

⚠️ Considerations

1. Hardware Requirements:

   • This workflow involves intensive video decoding and reconstruction, requiring a GPU with at least 16GB VRAM for optimal performance.

2. Resolution Limitations:

   • High-resolution input videos (above 1080p) may cause memory overflow.

   • To avoid crashes, keep the resolution at or below 1080p.

3. Model Compatibility:

   • Ensure that the model and ComfyUI versions are compatible to prevent rendering errors.

4. Output Quality Control:

   • Use negative prompts to reduce artifacts, blurriness, and inconsistent frame details.