Fooocus-Pro v0.1.0 — Experimental DirectML & Intel GPU Support

Fooocus-Pro v0.1.0

First public release of Fooocus-Pro.

Highlights

• Enhanced Fooocus inference pipeline
• Adaptive CFG support
• Adaptive anisotropic processing
• ADM conditioning improvements
• ControlNet softness adjustment
• Improved timestep/diffusion progress handling
• LoRA / LoKr / LoHa / GLora patch handling
• DirectML compatibility work for Intel GPUs
• DirectML-aware Brownian noise sampling
• Low-VRAM model management improvements
• Precision and CLIP patch integration

Hardware

Tested development environment:

• Windows 64-bit
• Python 3.11.6
• PyTorch 2.4.1
• Intel Iris Xe Graphics
• Intel graphics driver 32.0.101.7026
• torch-directml 0.2.5.dev240914

Experimental DirectML Support

Fooocus-Pro can initialize and execute the SDXL diffusion pipeline through
Microsoft DirectML on Intel graphics hardware.

The current Intel Iris Xe testing successfully completed the SDXL sampling
stage, but high-resolution VAE decoding may exceed available GPU resources
and cause a DirectML device reset.

DirectML support should therefore be considered experimental in v0.1.0.

Known Limitations

• Intel integrated GPUs have limited graphics memory.
• SDXL generation can be extremely slow on integrated graphics.
• High-resolution VAE decoding may fail due to GPU memory/device suspension.
• NVIDIA CUDA remains the preferred configuration for high-performance
  generation.

Status

Experimental / early public release.

More hardware compatibility testing and VAE memory optimization are planned
for future releases.