1. Workflow Overview

This is an anime-style video generation workflow based on WAN2.1 model (DaKai optimized), featuring:

• Convert input video (e.g., dance.mp4) to anime style

• Dynamic prompts for character details (19yo Chinese schoolgirl dancing)

• HunyuanLoom technology for motion coherence

• Outputs 16fps MP4 video (H.264 encoded)

2. Core Models

3. Key Components

Special Nodes:

• VHS_LoadVideo: Frame extraction (skips 120 frames, keeps 81)

• HYFlowEditGuiderCFG: Dynamic CFG guidance (CFG=7.5)

• SamplerCustomAdvanced: Advanced sampler (16 steps, simple scheduler)

• VAEDecodeTiled: Tiled decoding (512x512 tiles, 64px overlap)

Installation:

1. Video Helper Suite: Install via ComfyUI Manager

2. HunyuanLoom: Manual install from GitHub

3. WAN2.1 Models: Download separately to models folder

4. Workflow Structure

Group 1: Load Models

• Load UNET/CLIP/VAE (wan2.1_t2v_1.3B_fp16 + umt5_xxl_fp16)

• Apply ApplyTeaCachePatch (strength=0.1) for acceleration

Group 2: Prompts

• Positive Prompt: Detailed character/scene description (amber eyes, JK uniform, beach sunset)

• Negative Prompt: 60+ filters for realism/low-quality

• FluxGuidance boosts prompt weights

Group 3: Video Input

• Resize input to 832x480 (nearest-exact)

• Tiled VAE encoding

Group 4: Sampling & Output

• Generate latent with HYFlowEditSampler (seed=123478)

• Tiled decode + video render (CRF=19, yuv420p)

5. Inputs & Outputs

Inputs:

• Video file (default: dance.mp4)

• Pre-configured prompts

• Frame rate (16fps)

Outputs:

• MP4 video (e.g., hyloom_00003.mp4)

• Optional intermediate frames

6. Notes

• VRAM: 16GB+ recommended (1.3B model is VRAM-heavy)

• Video Specs: Input ≥1280x720, ≤30sec duration

• Motion Tuning: Adjust motion_coherence in HYFlowEditGuiderCFG if flickering occurs

• Common Error: Missing umt5_xxl_fp16 breaks CLIP encoding