Questions

Docs · Questions

StreamDiffusionTDUsage

Which GPU and settings should I use for StreamDiffusionTD?

Choose a starting configuration

For an 8 GB GPU such as the RTX 3070, start with stabilityai/sd-turbo at 512 x 512, with TensorRT enabled and optional features off.

Choose by dedicated GPU memory (VRAM). GPU series alone is not enough: cards in the same series can have different memory capacities.

Suggested starting points

These are conservative suggestions, not a tested compatibility chart or a guarantee that every configuration will fit.

Dedicated VRAM Model to start with Resolution Features for the first run
6-8 GB stabilityai/sd-turbo 512 x 512 TensorRT on; ControlNet, IP Adapter, Cached Attention and Feature Injection off
10-12 GB stabilityai/sd-turbo first; try SDXL-Turbo after a successful baseline 512 x 512 Keep optional features off while checking available memory
16 GB Try stabilityai/sdxl-turbo 512 x 512 Start with optional features off; add one at a time
24 GB or more Try stabilityai/sdxl-turbo 512 x 512 Establish a baseline, then add ControlNet, IP Adapter or temporal features one at a time

Use TensorRT for these starting configurations. Leave Tiny VAE enabled for the initial run. Higher resolution and additional features can increase memory use. TouchDesigner and other GPU applications also need room. More VRAM does not guarantee every feature combination will fit or a particular frame rate.

First run on an RTX 3070 / 8 GB GPU

  1. Stop the stream if it is running.
  2. Set Model Id to stabilityai/sd-turbo.
  3. Set the output size to 512 x 512 and Acceleration to tensorrt.
  4. Turn ControlNet, IP Adapter, Use Cached Attention (V2V) and Enable Feature Injection (V2V Smoothness) off. Leave other settings at their defaults for this first check.
  5. Start the stream. The first run may download the model and build its TensorRT engines.

This is a starting configuration to try, not a verified fix for every startup error. IP Adapter is not compatible with sd-turbo; use a supported SDXL or SD 1.5 configuration if you need it.

Dedicated VRAM, shared GPU memory and system RAM

  • Dedicated GPU memory is the VRAM used for the table above.
  • Shared GPU memory comes from system RAM. A display of 8 GB dedicated plus 16 GB shared does not make the card a 24 GB VRAM GPU.
  • System RAM is separate. Model loading and the first engine build can need more host memory than normal streaming. Having 32 GB of system RAM does not establish that a startup error is caused by insufficient RAM.

SDXL startup error (Failed to serialize proto)

A Failed to serialize proto error in SDXL TensorRT preparation came from the model-size check before the engine is built. It does not mean a wrong install, too little VRAM or a wrong protobuf version. The fix is in the current core: update to v0.4.2 (About page, then Install Step 1, Update). Keep your existing install and engine files. If the error continues, share the full console traceback, operator version, GPU and system RAM in support.

See the installation guide for setup and the parameter reference for controls.