Docs · Questions
StreamDiffusionTDUsage
What are all the StreamDiffusionTD parameters and what do they do?
StreamDiffusionTD Parameters Reference
StreamDiffusionTD has 100+ parameters across 9 pages. For the full reference, see the Parameters documentation.
This FAQ covers the highlights and what changed in v0.3.1.
New in v0.3.1
StreamV2V / Cached Attention (Settings 2)
- Cattenable: Toggle cached attention for video-to-video mode (requires TRT)
- Cattmaxframes: Number of frames in the attention cache
- Cattinterval: How often the cache refreshes
Scheduler & Sampler (Settings 2)
- Scheduler: Select noise scheduler (EulerDiscrete, TCD, LCM, etc.)
- Sampler: Select sampler algorithm
FX Processors (FX Processors page)
- Add and configure image-space effects that run in the diffusion feedback loop
- Ships with Feedback Loop (zoom, pan, rotation) and Feedback Grade (color grading, hue drift)
- Use the + button to add processors, parameters update live
Other New Parameters
- Similar_filter_sleep_fraction: Controls GPU idle time when frames are similar (0.0 to 1.0)
- Benchmark: Pulse button that runs a benchmark report and copies results to clipboard
- Usertox / Devtox mode: Toggle between User and Dev tox mode to preserve internal operator modifications across updates
Core Parameters (All Versions)
Settings 1 - Generation
- Modelid: HuggingFace ID or local path (default:
stabilityai/sd-turbo) - Width/Height: Output resolution (default: 512x512)
- Promptdict: Expandable prompt blocks with weight control
- Tindexblock: Denoising steps (lower = more random, higher = closer to input)
- Seeddict: Multiple seeds with weights for blending
- Ipadapterenable: Toggle IP Adapter (must be ON at engine build time)
- Guidancescale: CFG scale (default: 1.0 for turbo models)
Settings 2 - Stream Control
- Sdmode:
img2imgortxt2img - Limitfps: Max FPS cap
- Pausestream: Pause/resume without stopping
ControlNet
- Usecontrolnet: Must be ON when starting stream
- Cn[n]id / Cn[n]preprocessor / Cn[n]weight: Per-slot ControlNet config
- Autopreprocess: Auto-select preprocessor based on model
Models
- Acceleration:
none,xformers,tensorrt - Customlcm: LCM-LoRA selection
- Uselora / Loradictblock: LoRA loading with weight control
Key Parameter Interactions
- IP Adapter + TensorRT: Must decide BEFORE engine build
- ControlNet + Engine: Toggling ControlNet may require engine rebuild
- Resolution + TensorRT: Can change freely within 384-1024px range
- Steps + TensorRT: Can switch between 1-4 steps without rebuild
- StreamV2V + TensorRT: Engines are locked to their build resolution
Full reference: dotsimulate.com/docs/streamdiffusiontd/parameters Version: v0.3.1 (March 2026)