Legacy versions

Current release is v0.4.1. This page keeps the older install and v0.2.99 V2V notes so they are not lost under the new docs.

Changelogs for v0.3.1 and v0.3.0 stay on What’s New.

Feature	v0.2.99	v0.3.0	v0.3.1	v0.4.0
SDXL	No	Yes	Yes	Yes
IP Adapter / FaceID	No	Yes	Yes	Yes, together with CN + V2V + live LoRA
StreamV2V	Yes (no TRT)	No	Yes (TRT)	Yes (TRT + Feature Injection)
FX Processors	No	Basic	2 built-in + custom	Same, feedback_loop reworked
TensorRT	Cloud only	Local + Cloud	Local + Cloud	10.16, CUDA 12.8 only
Installer CLI	No	No	verify / repair / diagnose	Full Update (fresh venv)
Default model	sd-turbo	sdxl-turbo	sdxl-turbo	sdxl-turbo

v0.3.1 installation

Use this only if you are staying on v0.3.1. For v0.4.0 see the Installation Guide.

v0.3.1 must be installed in a new folder, separate from v0.2.99 or v0.3.0. Do not name the folder streamdiffusion.

Prerequisites

Python 3.11.9 (3.10.9 also worked on v0.3.1). Not 3.12, not 3.13. Add to PATH
Git, add to PATH
CUDA 12.8 recommended

Local steps

Drag the v0.3.1 tox into TouchDesigner
Install page: set Basefolder to a new folder
Clone StreamDiffusion (Step 1)
Select CUDA 12.8
Install StreamDiffusion (venv)
Install TensorRT (required for ControlNet, IP Adapter, StreamV2V)
Activate venv, then:

cd StreamDiffusion-installer
python -m sd_installer verify

If anything fails: python -m sd_installer repair. For bug reports: python -m sd_installer diagnose.

First run (v0.3.1)

Acceleration = tensorrt. First engine build takes a while. Suggested test: stabilityai/sd-turbo, 512x512, IP Adapter off, ControlNet weight 0.

Upgrading to v0.3.1 (historical)

From v0.3.0: load the v0.3.1 tox, Install Step 1 (should show X), fetch the pinned branch, then verify
From v0.2.99 or earlier: fresh folder, do not upgrade in place

Daydream on v0.3.1

Drag and drop. Select Daydream on the Install page. Hit the Daydream Login pulse. There is no API key to paste. Pinned to Daydream v0.2.2.

v0.3.0

November 6, 2025 rebuild: SDXL, IP Adapter, local TensorRT, Daydream drag-and-drop. No StreamV2V. Changelog is on What’s New. If you are still on v0.3.0, install v0.4.0 in a new folder via Full Update / Fresh Install rather than trying to salvage this tree.

v0.2.99

v0.2.99 is the last version with StreamV2V that does not need TensorRT. Keep it only if you need that path.

When v0.2.99 may still be useful:

V2V without TensorRT
A simpler setup that already works
Limited VRAM, sd-turbo default

V2V Temporal Consistency (v0.2.99)

V2V (Video-to-Video) in v0.2.99 provides smooth frame transitions using a different approach than v0.3.1’s cached attention. This version does NOT require TensorRT.

V2V Parameters

Parameter	Description	Default
Enable V2V	Enable temporal consistency	Off
Cache Attn Active	Store attention maps from past frames	On
Cache Interval	How often to update feature bank	1
Cache Max Frames	Number of frames stored	3
Use Feature Injection	Incorporate past frame details	On
Feature Injection Strength	Influence of past frames (0-1)	0.7
Feature Similarity Threshold	Minimum similarity for reuse	0.3
Use Tome Cache	Optimized cache storage	On
Tome Ratio	Cache allocation ratio	0.5
Use Grid	Grid-based feature organization	On

V2V Detailed Descriptions

Cache Attn Active: Enables the model to store attention maps from past frames, reusing them for processing new frames. Helps maintain consistency and reduces flicker.

Cache Interval: Sets how often the feature bank updates with new frames. Lower values mean less frequent updates, capturing fewer changes and reducing computational load.

Cache Max Frames: Limits the number of frames stored in the feature bank. Higher values store more frames, enhancing temporal consistency but using more memory.

Use Feature Injection: Incorporates details from past frames into the current frame’s processing. Helps maintain continuity and reduces flicker.

Feature Injection Strength: Controls how much influence past frame details have on the current frame. Higher values increase consistency but may introduce artifacts.

Feature Similarity Threshold: Sets the minimum similarity score for features from past frames to be reused. Uses cosine similarity, only reusing features above this threshold.

Use Tome Cache: Activates an optimized cache for storing features efficiently. Reduces memory usage and enhances real-time processing performance.

Tome Ratio: Defines the proportion of total cache allocated to Tome Cache. Higher ratios optimize for speed at the cost of memory.

Use Grid: Activates a grid-based approach for organizing and processing features. Divides the frame into a grid, improving matching and fusion of features from past frames.

V2V Important Notes

Not compatible with TensorRT acceleration (unlike v0.3.1’s cached attention which requires TRT)
Increases VRAM usage
Best for video sequences where frame-to-frame smoothness matters
May reduce FPS compared to non-V2V mode

v0.2.99 Installation

Prerequisites

Python 3.11.9 or 3.10.9
CUDA 12.1 (or 12.8 for RTX 50-series)
Git

Installation Steps

Load operator: Drag StreamDiffusionTD.tox into TouchDesigner
Go to Install page
Download StreamDiffusion to a folder
Install Virtual Environment
(Optional) Install TensorRT

Upgrading

If upgrading from earlier versions:

v0.2.6+: Automatically detects previous installation
Earlier: Set Basefolder parameter manually

v0.2.99 System Requirements

Local Mode

OS: Windows 10/11, macOS (Apple Silicon limited)
GPU: NVIDIA with 6GB+ VRAM (RTX 3060+)
CUDA: 12.1 recommended (12.8 for RTX 50-series)
Python: 3.11.9 or 3.10.9

v0.2.99 Key Features

sd-turbo as default model
Daydream cloud backend support
RTX 5090/5080 compatibility (limited)
Enhanced Mac installation (cloud mode)
Feedback loop toggle
V2V temporal consistency (works without TensorRT)

Running Both Versions

You can run both versions:

Install v0.2.99 in one folder (e.g., C:/AI/StreamDiff_v0299/)
Install v0.3.1 in another folder (e.g., C:/AI/StreamDiff_v031/)
Use different TOX files pointing to different base folders
Switch between them based on your needs

Download

v0.2.99 is available on Patreon alongside v0.3.1.