fblissjr-ComfyUI-WanActivationEditor/docs/ACTIVATION_USER_GUIDE.md

WanVideo Activation Editor User Guide

Overview

Block-level prompt injection for WanVideo models. Inject alternative text conditioning into specific transformer blocks to blend different concepts at various processing depths.

• What it does: Mixes different prompts at specific transformer blocks
• What it doesn't do: Change prompts over time (not temporal prompt travel)
• Key mechanism: 40 transformer blocks, each can use different text embeddings

How It Works

1. Text → T5 encoder → embeddings (4096-dim)
2. Projection → transformer space (5120-dim)
3. 40 transformer blocks process sequentially
4. Each block can use different text embeddings via cross-attention
5. Blended embeddings guide generation

The editor stores configuration in transformer_options for WanVideoWrapper to process during generation.

Quick Start

1. Two WanVideoTextEncode nodes: main prompt + injection prompt
2. WanVideoBlockActivationBuilder: select preset or custom blocks
3. WanVideoActivationEditor: set strength (0.3-0.5) and log level
4. Continue normal generation

Critical: Prompts must be >50% different for visible effects!

Example:

Main: "A gray tabby cat walks through a sunlit garden. The cat moves gracefully between 
      rows of colorful flowers, its tail swaying gently. Butterflies flutter nearby."
      
Injection: "An underwater coral reef scene with crystal clear water. Schools of tropical 
           fish swim between vibrant coral formations. Sunlight filters through the water."
           
Blocks: late_blocks
Strength: 0.4
Result: Cat with aquatic features, garden with underwater lighting effects

Block Patterns

General Tendencies (not rigid rules)

• 0-10: Often affects colors, textures, low-level features
• 10-20: May influence shapes, local structures
• 20-30: Can affect objects, composition
• 30-40: Typically handles semantic meaning

Presets

• early_blocks (0-10): Texture/color experiments
• mid_blocks (10-30): Structure blending
• late_blocks (30-40): Semantic mixing
• alternating: Every 2nd block
• sparse: Every 4th block
• first_half (0-20): Foundation layer
• second_half (20-40): Higher concepts

Workflow Setup

1. Load model with WanVideoModelLoader
2. Encode main and injection prompts with WanVideoTextEncode
3. Select blocks with WanVideoBlockActivationBuilder
4. Apply with WanVideoActivationEditor (strength 0.3-0.5)
5. Generate with WanVideoSampler

Lightx2v Notes

• 4-step inference amplifies effects
• Use lower strength (0.1-0.2)
• Late blocks work best

Test Examples (WAN-Style Prompts)

Environmental Transformation

Office → Volcanic (late blocks, strength 0.4)

• Main: "A woman sits at a wooden desk in a modern office, typing on a silver laptop. Behind her, floor-to-ceiling windows reveal a cityscape. The office has minimalist decor with white walls."
• Injection: "Molten lava flows down a volcanic mountainside. The glowing orange-red magma radiates intense heat, causing the air to shimmer. Smoke and steam rise from the flow."
• Result: Office with lava-like lighting, heat distortion effects

Living Room → Blizzard (mid blocks, strength 0.5)

• Main: "A family gathers around a stone fireplace in a cozy living room. Children play on the carpet while parents relax on a leather sofa. Warm lighting from table lamps creates intimacy."
• Injection: "A powerful blizzard engulfs a mountain landscape. Snow falls heavily, driven by fierce winds. Visibility is reduced to mere meters as the storm rages."
• Result: Interior with frost patterns, cold blue lighting

Activity Contrast

Gardening → DJ Performance (alternating blocks, strength 0.3)

• Main: "An elderly man tends to his garden in morning light. He kneels beside tomato plants, pruning with small shears. His weathered hands move with practiced precision."
• Injection: "A DJ performs at a packed nightclub. Hands move rapidly across turntables. Strobe lights flash in sync with pounding bass. The crowd dances energetically."
• Result: Surreal mix of peaceful gardening with rhythmic energy

Material Transformation

Tech Lab → Rainforest (early blocks, strength 0.6)

• Main: "A robotics laboratory buzzes with activity. Robotic arms perform assembly tasks. Engineers monitor screens showing data. The space has clean white surfaces."
• Injection: "An ancient rainforest teems with life. Massive trees rise like pillars, covered in moss and vines. Exotic birds call from the canopy. Mist hangs in the air."
• Result: Lab equipment with organic, moss-like textures

Custom Patterns

• Bookend: 1111111111000000000000000000001111111111 (strong start/end)
• Wave: 0011110000111100001111000011110000111100 (periodic influence)

Advanced Techniques

• Multi-stage: Chain editors for layered effects
• Strength ranges: 0.1-0.3 (subtle), 0.4-0.6 (balanced), 0.7-1.0 (strong)
• Complementary prompts: Work best when sharing conceptual elements
• Compatible with: NAG, Radial Attention, LoRA

Troubleshooting

No visible effect:

• Check prompts are >50% different (see debug output)
• Ensure some blocks are activated (not all zeros)
• Try increasing strength to 0.6-0.8
• Verify all connections are made

Too strong/chaotic:

• Reduce strength to 0.1-0.3
• Use fewer active blocks
• Try sparse preset instead of continuous blocks

Checking prompt difference: Set log level to "verbose" and look for:

[WAN_DEBUG] Block 0 Blending Results:
  - Percent changed: 78.4%  ✓ GOOD - very different
  - Percent changed: 1.1%   ✗ BAD - too similar

Verifying activation:

[WanVideoActivationEditor] Active blocks: [30, 31, 32, 33, 34, 35, 36, 37, 38, 39]
✓ Model patched successfully - 10 blocks will be modified

Best Practices

WAN Prompt Style

Write detailed scene descriptions including:

• Subject appearance and actions
• Environment and setting details
• Lighting and atmosphere
• Object positions and relationships
• Colors, textures, materials

Effective Workflow

1. Test prompt difference first: Use WanVideoVectorDifference to verify >50% difference
2. Start with presets: early_blocks for textures, late_blocks for concepts
3. Initial strength: 0.3-0.5 (reduce to 0.1-0.2 for Lightx2v)
4. Use log level: Set to "basic" or "verbose" to monitor activation

Performance

• No additional VRAM usage (model not duplicated)
• Minimal computation overhead
• Works with all WanVideo model variants

Creative Applications

• Style transfer: Early blocks
• Object morphing: Mid blocks
• Concept blending: Late blocks
• Mood shifts: All blocks, low strength

Example Results

Underwater Office (late_blocks, 0.4): Office with aquatic lighting, floating papers Mechanical Garden (mid_blocks, 0.5): Flowers with metallic petals Time Blend (custom middle blocks, 0.6): Medieval castle with futuristic elements