TouchDesigner Features
The Daydream TouchDesigner plugin exposes all the power of StreamDiffusion through an intuitive parameter interface.
Parameters Reference
| Parameter | Description |
|---|
| Prompt Schedule | Weighted prompts: [("anime style", 1.0), ("oil painting", 0.5)] |
| Prompt Interpolation | How to blend between prompts: linear or slerp |
| Negative Prompt | What to avoid: “blurry, low quality, flat” |
| Seed Schedule | Weighted seeds: [(42, 1.0), (123, 0.5)] |
| Seed Interpolation | How to blend between seeds: linear or slerp |
| Randomize Seeds | Button to set all seeds to random values |
| Guidance | How closely to follow the prompt (1.0-3.0) |
| Delta | Strength of diffusion effect (0.0-1.0) |
| Steps | Number of inference steps (1-4) |
| ControlNet scales | Strength of each conditioning type (0.0-1.0) |
| IP Adapter | Enable/disable style transfer |
| IP Adapter Scale | Strength of style influence |
| IP Adapter Type | regular or faceid |
Prompt Scheduling
The killer feature for live performances: smoothly transition between prompts over time.
# Single prompt
prompt = "cyberpunk cityscape, neon lights"
# Weighted blend of prompts
prompt_schedule = [
("anime style portrait", 0.7),
("oil painting", 0.3),
]
slerp interpolation, transitions are smooth and organic - perfect for VJ sets.
Use prompt scheduling with TouchDesigner’s animation channels to create evolving visuals that respond to music or time.
Seed Scheduling
Control randomness while maintaining smooth transitions:
# Single seed for consistency
seed = 42
# Weighted blend of seeds
seed_schedule = [
(42, 0.8),
(123, 0.2),
]
ControlNets
ControlNets preserve structure from your input video. Available types depend on the model:
SDXL Models
| ControlNet | What It Preserves | Best For |
|---|
| Depth | 3D structure, distance | Faces, scenes with depth |
| Canny | Sharp edges, outlines | Line art, detailed edges |
| Tile | Texture patterns | Upscaling, detail preservation |
SD1.5 Models
| ControlNet | What It Preserves | Best For |
|---|
| Depth | 3D structure | Faces, objects |
| Canny | Sharp edges | Linework, silhouettes |
| Tile | Texture | Detail preservation |
| TemporalNet | Motion consistency | Reducing flicker |
SD2.1 Models
| ControlNet | What It Preserves | Best For |
|---|
| OpenPose | Body poses | Dance, movement |
| HED | Soft edges | Organic shapes |
| Canny | Sharp edges | Architecture |
| Depth | Distance | Spatial scenes |
| Color | Color composition | Palette preservation |
| TemporalNet | Motion | Video stability |
Adjust ControlNet scales to balance between following the prompt and preserving input structure. Start with 0.5 and adjust from there.
IP Adapter (Style Transfer)
Apply the style of a reference image to your video:
- Enable IP Adapter in the parameters
- Load a style reference image
- Adjust Scale (higher = stronger style influence)
- Choose Type:
regular - General style transferfaceid - Face-specific (SDXL only)
IP Adapter works best with SDXL and SD1.5 models. SD2.1 doesn’t support it.
Model Selection
Choose the right model for your use case:
| Model | Speed | Quality | ControlNets | IP Adapter |
|---|
| SD Turbo | Fastest | Good | 6 types | No |
| SDXL Turbo | Fast | Best | 3 types | Yes |
| Dreamshaper 8 | Medium | Great | 4 types | Yes |
| Openjourney v4 | Medium | Great | 4 types | Yes |
- Resolution: Start with 512x512, increase only if needed
- Steps: 2-3 steps is usually enough for real-time
- ControlNets: Enable only what you need; each adds latency
- Delta: Lower values = more stable, higher = more responsive
Example: VJ Setup
Here’s a typical setup for a live VJ performance:
- Input: Live camera or video loop
- Model: SDXL Turbo for quality, SD Turbo for speed
- ControlNets:
- Depth at 0.6 (preserve structure)
- Canny at 0.3 (keep edges)
- Prompt Schedule: Animate between themes with CHOPs
- IP Adapter: Cycle through style images
Next Steps
