LTX-2, Flux2 Klein & GGUF Support
This update adds support for LTX-2, Flux2 Klein, and GGUF checkpoints. Optimized for performance, new FP8 quantization allows for high-quality inference on a wider range of hardware
What’s New
New Pipelines
- LTX-2
- Flux2 Klein
- Z-Image ControlNet
New Features
- Support for GGUF diffusers checkpoints.
- Support for float8 quantization (CUDA only)
Bug Fixes & Stability Improvements
- Various fixes and internal cleanups to improve reliability during model loading and inference.
Full Changelog: v0.5.1...v0.5.5
Installation
1. Installer Version
- Uninstall Diffuse vX.X.X
- Download and run Diffuse_v0.5.5.exe
- Follow the on-screen instructions
2. Standalone Version
-
Download and extract Diffuse_v0.5.5.zip
A fast SSD with plenty of free space is recommended, as model downloads can be large. -
Run Diffuse.exe
Diffuse will automatically:
- Install an isolated portable Python runtime
- Create the required virtual environment
- Download the selected model from Hugging Face
First-run notice
On first launch or when loading a model for the first time, setup may take several minutes while Python, dependencies, and model files are downloaded and initialized. This is expected behavior.
No manual Python setup is required.
Important (Beta Notice)
Diffuse is still in Beta, and during this phase some releases include large internal changes.
For this version, a full uninstall and reinstall is recommended to avoid upgrade issues.
Your downloaded models are not affected and can be kept.
Your image/video history are not affected and can be kept.
Settings.json will be replaced during installation
If you have added custom environments, paths, or models, please back up your Settings.json file before upgrading.
These upgrade steps are temporary and expected during Beta. The goal is to stabilize updates and avoid full re-installs for future releases.
Device Support
Full Support for CUDA based devices.
Experimental support for ROCM based devices.
⚠️ ROCM Compatibility Note:
Experimental support is active, but a known issue with VAE tiling on AMD hardware causes unrecoverable system freezes for LTX-2, WAN, and Qwen. We recommend avoiding these models on ROCm until a driver-level fix is available.
LoRA support is currently disabled for ROCm devices due to the lack of torch.distributed support in the standard ROCm PyTorch binaries.
