Parameters Reference (v0.3.1)

StreamDiffusionTD has 100+ parameters organized across multiple pages. This reference covers the most important parameters.

Page Navigation

Settings 1 - Core generation, prompts, seeds, IP Adapter
Settings 2 - Stream controls, connection settings, StreamV2V
Settings 3 - Configurations, presets, TensorRT engines
FX Processors - Image/latent processing hooks
ControlNet - Image conditioning controls
Models - Model loading, LoRA, VAE, scheduler/sampler
Install - Installation and backend configuration
About - Version info, updates, benchmark
Callbacks - TouchDesigner callback functions

Settings 1

Core generation parameters including prompts, seeds, IP Adapter, and guidance.

Model ID (Modelid) op('StreamDiffusionTD').par.Modelid StrMenu

The model to use for generation. Default is SDXL-Turbo.

Default: stabilityai/sdxl-turbo Options: stabilityai/sd-turbo, stabilityai/sdxl-turbo, prompthero/openjourney-v4

Width / Height (Width, Height) op('StreamDiffusionTD').par.Width Int

Output resolution. TensorRT supports 384-1024px without engine rebuild.

Default: 512 Range: 384-1024 (auto-rounds to multiple of 64)

Prompts

Prompt Blocks (Promptdict) op('StreamDiffusionTD').par.Promptdict Sequence

Multiple prompts with individual weights for blending. Use +/- buttons to add/remove prompts.

Default: "default fancy banana" Sequence Block with concept and weight

Interpolation (Setinterpolation) op('StreamDiffusionTD').par.Setinterpolation Menu

Method for interpolating between multiple prompts. "average" = linear blend of embeddings. "slerp" = spherical interpolation for smoother morphing between prompts.

Default: slerp Options: average, slerp

Step Schedule (T Index)

Step Schedule (Tindexblock) op('StreamDiffusionTD').par.Tindexblock Sequence

Similar to 'denoise' in img2img. Each step has a slider (1-49). Lower values = more randomness, higher values = closer to input. Adjustable while streaming. Step count can change without TensorRT rebuild.

Default: 35 Range: 1-49 per step

Seeds

Seed Blocks (Seeddict) op('StreamDiffusionTD').par.Seeddict Sequence

Multiple seeds with weights for blending. Use -1 for random seed every frame (creates variation/flickering effect).

Default: Random Sequence Block with seed value and weight

IP Adapter

IP Adapter Enable (Ipadapterenable) op('StreamDiffusionTD').par.Ipadapterenable Toggle

Enable IP Adapter for image prompt conditioning. Must be decided BEFORE engine build. Cannot toggle after.

Default: On Options: On | Off

IP Adapter Scale (Ipadapterscale) op('StreamDiffusionTD').par.Ipadapterscale Float

Strength of IP Adapter influence on generation.

Default: 0.0 Range: 0.0-1.0

IP Adapter Image (Ipadapterimage) op('StreamDiffusionTD').par.Ipadapterimage TOP

Reference image TOP for IP Adapter conditioning.

Default: None TOP reference

FaceID Mode (Ipfaceid) op('StreamDiffusionTD').par.Ipfaceid Toggle

Enable face-specific IP Adapter mode for face preservation.

Default: Off Options: On | Off

Guidance

Guidance Scale (Guidancescale) op('StreamDiffusionTD').par.Guidancescale Float

CFG scale controlling adherence to prompt.

Default: 1.0 Range: 0.0+

Delta (Delta) op('StreamDiffusionTD').par.Delta Float

Delta multiplier of virtual residual noise.

Default: 1.0 Range: 0.0+

CFG Type (Cfgtype) op('StreamDiffusionTD').par.Cfgtype Menu

CFG mode for img2img. Negative prompts only work with "full" mode (results not ideal).

Default: self Options: none, full, self, initialize

Settings 2

Stream control, connection settings, mode selection, and StreamV2V.

SD Mode (Sdmode) op('StreamDiffusionTD').par.Sdmode Menu

Switch between image-to-image and text-to-image generation modes.

Default: img2img Options: img2img, txt2img

Pause Stream (Pausestream) op('StreamDiffusionTD').par.Pausestream Toggle

Pause/resume streaming without stopping the server.

Default: Off Options: On | Off

Debug Mode (Debugmode) op('StreamDiffusionTD').par.Debugmode Toggle

Enable verbose debug logging for Local backend (td_main.py).

Default: On Options: On | Off

Limit FPS (Limitfps) op('StreamDiffusionTD').par.Limitfps Int

Maximum frames per second cap. Useful for controlling GPU load. Applies instantly without restart.

Default: 60 Range: 1-120

Similar Filter Sleep (Similarfiltersleepfraction) op('StreamDiffusionTD').par.Similarfiltersleepfraction Float

When the similar image filter skips a frame (because the input hasn't changed enough), this controls how long the GPU sleeps. Prevents the GPU from spinning at 100% on static input without adding visible latency.

Default: 0.025 Range: 0.0-1.0

StreamV2V / Cached Attention (New in v0.3.1)

Video-to-video mode using cached attention maps for temporal consistency between frames. Requires TensorRT acceleration and the peft package (installed automatically).

Cached Attention Enable (Cattenable) op('StreamDiffusionTD').par.Cattenable Toggle

Enable StreamV2V cached attention for temporal consistency. Must be set before starting the stream.

Default: Off Options: On | Off

Cached Attention Max Frames (Cattmaxframes) op('StreamDiffusionTD').par.Cattmaxframes Int

Number of frames stored in the attention cache. Higher values increase temporal consistency but use more VRAM.

Default: 3 Range: 1-10

Cached Attention Interval (Cattinterval) op('StreamDiffusionTD').par.Cattinterval Int

How often the attention cache updates with new frames. Lower values update more frequently.

Default: 1 Range: 1-10

StreamV2V Notes:

Requires TensorRT acceleration (non-TRT paths are broken for this feature)
Resolution is locked to the TRT engine build dimensions (e.g., 512x512 engine only works at 512x512)
Uses a circular buffer for the KVO cache, optimized for performance

FX Processors

Hook system for custom image and latent processing at four pipeline stages. See the FX Processors page for full documentation on each processor and its parameters.

FX Processors (Fx) op('StreamDiffusionTD').par.Fx Sequence

Add FX processors for image/latent pre and post processing. Parameters for each processor are generated dynamically on the operator when added.

Default: None Sequence Block

Built-in Processors

Processor	Stage	Description
feedback_loop	image_pre	Deforum-style feedback with zoom, pan, rotation
feedback_grade	image_pre	Color grading that compounds through feedback (brightness, contrast, saturation, gamma, hue, temperature, invert)

Processing Stages

FX Processors can be applied at four stages:

Image Preprocessing (image_pre): Before VAE encoding. Both built-in processors run here so their effects compound through the feedback loop
Latent Preprocessing (latent_pre): After encoding, before diffusion (open for custom processors)
Latent Postprocessing (latent_post): After diffusion, before decoding (open for custom processors)
Image Postprocessing (image_post): After decoding (open for custom processors)

ControlNet

Image conditioning controls for both local and Daydream modes.

Use ControlNet (Usecontrolnet) op('StreamDiffusionTD').par.Usecontrolnet Toggle

Enable ControlNet conditioning. Must be ON when starting stream.

Default: On Options: On | Off

ControlNet Blocks (Cn) op('StreamDiffusionTD').par.Cn Sequence

Multiple ControlNet configurations. Each has model ID, preprocessor, weight, and enable toggle. v0.3.1 supports verified dual ControlNet streaming.

Default: depth Sequence Block

Auto Preprocess (Autopreprocess) op('StreamDiffusionTD').par.Autopreprocess Toggle

Automatically select a preprocessor based on the ControlNet model. Turn this OFF if you want to set preprocessors manually, otherwise your selection will be overridden.

Default: On Options: On | Off

ControlNet Models (SDXL)

For SDXL-Turbo, use SDXL-compatible ControlNets:

xinsir/controlnet-depth-sdxl-1.0
xinsir/controlnet-canny-sdxl-1.0
xinsir/controlnet-tile-sdxl-1.0

Preprocessors

Available preprocessors:

canny - Edge detection
depth - Depth estimation (CPU, higher VRAM)
depth_tensorrt - Depth estimation (GPU accelerated, auto-builds TRT engine on first use, ~60% faster than CPU depth)
hed - Holistically-nested edge detection
external - Use pre-processed input
feedback - Use previous output as conditioning

Models

Model loading, acceleration, LoRA, VAE, and scheduler/sampler configuration.

Acceleration (Acceleration) op('StreamDiffusionTD').par.Acceleration Menu

Acceleration method. TensorRT is required for ControlNet, IPAdapter, and StreamV2V.

Default: tensorrt Options: none, xformers, tensorrt

Scheduler (Scheduler) op('StreamDiffusionTD').par.Scheduler Menu

Noise scheduler selection. TCD scheduler is required for StreamV2V / cached attention.

Default: lcm Options: lcm, tcd

Sampler (Sampler) op('StreamDiffusionTD').par.Sampler Menu

Sampling method for the diffusion process.

Default: simple Options: simple, sgm uniform, normal, ddim, beta, karras

Custom LCM (Customlcm) op('StreamDiffusionTD').par.Customlcm StrMenu

LCM-LoRA selection. Auto-selects based on model architecture.

Default: auto Options: auto, latent-consistency/lcm-lora-sdv1-5, latent-consistency/lcm-lora-sdxl

Custom VAE (Customvae) op('StreamDiffusionTD').par.Customvae StrMenu

VAE selection for encoding/decoding.

Default: auto Options: auto, madebyollin/taesd, madebyollin/taesdxl, stabilityai/sd-vae-ft-mse

Use LoRA (Uselora) op('StreamDiffusionTD').par.Uselora Toggle

Enable LoRA model loading. LoRA support is experimental/untested. Note: sd-turbo is SD 2.1 based, so SD 1.5 LoRAs won't work with it.

Default: Off Options: On | Off

Install

Backend and installation configuration.

Backend (Backend) op('StreamDiffusionTD').par.Backend Menu

Choose between local GPU processing or Daydream cloud.

Default: Daydream Options: local, Daydream

Base Folder (Basefolder) op('StreamDiffusionTD').par.Basefolder Folder

Location of StreamDiffusion installation.

Default: Empty Directory path

User/Dev Tox Mode (Loadusertox) op('StreamDiffusionTD').par.Loadusertox Toggle

Toggle between loading the verified operator TOX and a user-modified copy. When enabled, the operator keeps a separate user copy that preserves your internal modifications across backend switches and updates.

Default: Off Options: On | Off

About

Version info, update checking, and benchmark tools.

Generate Benchmark (Generatebenchmark) op('StreamDiffusionTD').par.Generatebenchmark Pulse

Gathers FPS, VRAM usage, current config, and GPU info and copies a formatted report to your clipboard. Useful for sharing performance results or reporting issues.

Check for Updates (Checkforupdates) op('StreamDiffusionTD').par.Checkforupdates Pulse

Check for new operator TOX versions. Requires sign-in on the About page. The operator checks automatically once every 24 hours if authorized.

Callbacks

TouchDesigner callback function controls.

Callback DAT (Callbackdat) op('StreamDiffusionTD').par.Callbackdat DAT

Reference to DAT containing callback functions.

Default: StreamDiffusionTD_callbacks DAT reference

Available Callbacks

onReceiveFrame: When new frame is received from StreamDiffusion
onStreamStart: When streaming starts
onStreamEnd: When streaming ends
onFrameReady: When frame is ready for processing
onImageChange: When input image changes

TensorRT Engine Behavior

TensorRT engines are flexible in v0.3.1:

Engine depends on:

Model
ControlNet enabled/disabled
IP Adapter enabled/disabled

Engine does NOT depend on:

Resolution (384-1024px, change freely)
Step count (1-4 steps, change freely)

When you must rebuild:

Changing model
Toggling IP Adapter
Toggling ControlNet

StreamV2V note: Cached attention engines are locked to their build resolution. If you built at 512x512, StreamV2V only works at 512x512. Change resolution requires an engine rebuild.