comfyui-video-pipeline

ComfyUI Video Pipeline

Orchestrates video generation across three engines, selecting the best one based on requirements and available resources.

Engine Selection

VIDEO REQUEST
    |
    |-- Need film-level quality?
    |   |-- Yes + 24GB+ VRAM → Wan 2.2 MoE 14B
    |   |-- Yes + 8GB VRAM → Wan 2.2 1.3B
    |
    |-- Need long video (>10 seconds)?
    |   |-- Yes → FramePack (60 seconds on 6GB)
    |
    |-- Need fast iteration?
    |   |-- Yes → AnimateDiff Lightning (4-8 steps)
    |
    |-- Need camera/motion control?
    |   |-- Yes → AnimateDiff V3 + Motion LoRAs

Pipeline 1: Wan 2.2 MoE (Highest Quality)

Image-to-Video

Prerequisites:

• wan2.1_i2v_720p_14b_bf16.safetensors in models/diffusion_models/
• umt5_xxl_fp8_e4m3fn_scaled.safetensors in models/clip/
• open_clip_vit_h_14.safetensors in models/clip_vision/
• wan_2.1_vae.safetensors in models/vae/

Settings:

Frame count guide:

VRAM optimization:

• FP8 quantization: halves VRAM with minimal quality loss
• SageAttention: faster attention computation
• Reduce frames if OOM

Text-to-Video

Same as I2V but uses wan2.1_t2v_14b_bf16.safetensors and EmptySD3LatentImage instead of image conditioning.

First+Last Frame Control (Wan 2.2 Exclusive)

Wan 2.2 MoE allows specifying both the first and last frame, enabling precise video planning:

1. Generate two hero images with consistent character
2. Use first as start frame, second as end frame
3. Wan interpolates the motion between them

Pipeline 2: FramePack (Long Videos, Low VRAM)

Key Innovation

VRAM usage is invariant to video length - generates 60-second videos at 30fps on just 6GB VRAM.

How it works:

• Dynamic context compression: 1536 markers for key frames, 192 for transitions
• Bidirectional memory with reverse generation prevents drift
• Frame-by-frame generation with context window

Settings

When to Use

• Videos longer than 10 seconds
• Limited VRAM systems (but RTX 5090 doesn't need this)
• When VRAM is needed for parallel operations
• Batch video generation

Pipeline 3: AnimateDiff V3 (Fast, Controllable)

Strengths

• Motion LoRAs for camera control (pan, zoom, tilt, roll)
• Effect LoRAs (shatter, smoke, explosion, liquid)
• Sliding context window for infinite length
• Very fast with Lightning model (4-8 steps)

Settings

Camera Motion LoRAs

Post-Processing Pipeline

After any video generation:

1. Frame Interpolation (RIFE)

Doubles or quadruples frame count for smoother motion:

Use rife47 or rife49 model.

2. Face Enhancement (if character video)

Apply FaceDetailer to each frame:

• denoise: 0.3-0.4 (lower than image - preserves temporal consistency)
• guide_size: 384 (speed optimization for video)
• detection_model: face_yolov8m.pt

3. Deflicker (if needed)

Reduces temporal inconsistencies between frames.

4. Color Correction

Maintain consistent color grading across frames.

5. Video Combine

Final output via VHS Video Combine:

Talking Head Pipeline

Complete pipeline for character dialogue:

Quality Checklist

Before marking video as complete:

• Character identity consistent across frames
• No flickering or temporal artifacts
• Motion looks natural (not jerky or frozen)
• Face enhancement applied if character video
• Frame rate is smooth (24+ fps for delivery)
• Audio synced (if talking head)
• Resolution matches delivery target

Reference

• references/workflows.md - Workflow templates for Wan and AnimateDiff
• references/models.md - Video model download links
• references/research-log.md - Latest video generation advances
• state/inventory.json - Available video models