What’s New

v0.3.1 - April 1, 2026

Stability and polish update following the v0.3.0 rebuild. Refined installation, new creative tools, and tons of fixes based on community feedback.

Since the March 24 early release on Discord: minor auto-update tweak and script version string bump.

Built-in FX Processors - Two processors ship with the operator, ready to use:
- Feedback Loop - Deforum-style image feedback with zoom, pan, rotation
- Feedback Grade - Color grading that compounds through the feedback loop (brightness, contrast, saturation, gamma, hue rotation, temperature, invert)
Custom FX Processor Support - Drop your own Python processors into custom_processors/ and they auto-load. Four pipeline stages available (image_pre, latent_pre, latent_post, image_post). Claude skill included to help you write them. See the FX Processors page for details
StreamV2V / Cached Attention - Video-to-video mode with configurable max frames and interval. Temporal consistency is back! Requires TensorRT
Depth TRT Auto-Build - Depth ControlNet TRT engine builds automatically on first use. Roughly 60% faster and a fraction of the VRAM compared to the PyTorch depth preprocessor
User/Dev Tox Mode - Toggle to preserve your internal operator modifications across backend switches and updates
FPS Limiting and Benchmark Report - Control GPU load, one-click benchmark to clipboard
Scheduler/Sampler Selection - Choose between LCM and TCD schedulers, plus multiple sampler options

New Installer CLI - verify, diagnose, and repair commands built into the operator
13-point verification covering PyTorch CUDA, StreamDiffusion core, numpy, diffusers fork, protobuf, onnx, peft, and more
Diagnose command outputs GPU info, full package versions, and environment details for bug reports
Repair command auto-fixes common issues (onnx version, diffusers fork, protobuf pin)
TensorRT install fixed on Windows - Installs sub-packages instead of the broken meta-package
Dependency conflicts resolved - Protobuf, numpy, onnx, and diffusers fork version conflicts are caught and fixed automatically
Commit tracking - Operator verifies your local repos match the expected version on startup
Installation debug Claude skill - Ships with the operator. If you use Claude Code, it knows how to diagnose and fix StreamDiffusionTD installs

Cloud operator is now a separate bundled component included in this release
ControlNet menu system works on cloud backend
Pinned to Daydream v0.2.2 (v0.2.3 had a stability issue, fix expected in v0.2.4)

TensorRT is required for ControlNet, IPAdapter, and StreamV2V (non-TRT paths are broken)
Dual ControlNet on 24GB GPUs runs near VRAM ceiling with reduced FPS
StreamV2V engines are locked to their build resolution
If you previously ran v0.3.0 dependency fix scripts, you may need to reinstall the diffusers fork (the repair command handles this)

Complete rebuild with SDXL support, IP Adapter, and TensorRT acceleration.

SDXL-Turbo as default model with native SDXL support
IP Adapter with FaceID for image prompt conditioning
Local TensorRT acceleration (previously cloud-only)
Daydream Cloud backend with zero installation
Multiple ControlNets simultaneously
FX Processor hook system for custom processing
Flexible TRT engines supporting 384-1024px without rebuild
New model compatibility system, YAML config generation, and SharedMemory backend communication

Bridge release between v0.2.x and the v0.3.0 rebuild.

SDXL ControlNet support, installation bug fixes, FPS counter

Resolution switching fix, improved local model and LoRA support

ControlNet support, V2V temporal consistency, Pause/Play/Unload commands, offline usage, LoRA improvements

Mac support, Model Preset Config, non-512x512 TensorRT resolutions

Major performance update: direct memory buffer replacing NDI+Spout

SDXL-Turbo and SDXS model support

Initial release: automated setup, stream control, live parameter updates, LoRA support, NDI/Spout integration