Version 5.0.0

Installation

Windows

1. Download dgenerate-installer-win-5.0.0.exe
2. Run the installer

Note for Windows users: Windows Defender may flag the installer as potentially harmful since it's unsigned. If you see a warning, click "More info" and then "Run anyway" to proceed with the installation.

Linux

1. Download dgenerate-installer-linux-5.0.0
2. Make the file executable: chmod +x dgenerate-installer-linux-5.0.0
3. Run the installer: ./dgenerate-installer-linux-5.0.0

macOS (Apple Silicon)

1. Download dgenerate-installer-macos-5.0.0
2. Open Terminal and navigate to Downloads: cd ~/Downloads
3. Remove quarantine attribute: xattr -d com.apple.quarantine dgenerate-installer-macos-5.0.0
4. Make the file executable: chmod +x dgenerate-installer-macos-5.0.0
5. Run the installer: ./dgenerate-installer-macos-5.0.0

Note for macOS users: The xattr command is required to prevent macOS from quarantining the binary, which would otherwise prevent execution.

---

Build Date: 2025-10-15 09:25:46 UTC
Workflow: Build and Publish Release #1

---

Patch Notes

• Python 3.13, torch 2.8.0
• Vastly improved shell syntax, no more !END statements, proper parsing of control structures
• Speed / resolution enhancement options: HiDiffusion, DeepCache, TeaCache, RAS, SADA, freeu etc.
• Prompt argument embedding (for certain arguments)
• LLM4GEN prompt weighter for SD1.5
• compel prompt weighter now supports Flux
• Local LLM based prompt upscaler plugins (GPT4ALL / transformers) and language translation via argostranslate or transformers (Helsinki NLP)
• Flux Kontext pipelines, SD3 Pix2Pix pipelines
• Enhanced offline mode for typical single file checkpoints (CivitAI), vendored HuggingFace configs used when --offline-mode active
• DynamicPrompts combinatoric prompting, and random attention (as prompt upscalers)
• Quantization support, backends bitsandbytes & sdnq (SD.Next quantization)
• Quantization support for single file checkpoints such as those from CivitAI, including with LoRAs
• Quantization specifiable on a per sub-model or global basis
• Quantization support for transformers based LLM prompt upscaler (souped up MagicPrompts)
• All applicable ControlNet union pipelines
• Automatic crop to inpaint area capability similar to Invoke
• All Kolors pipelines including new custom ControlNet pipelines
• Full scheduler argument parameterization from the command line / shell
• Partial denoising and latents interchange (latents as output / input)
• Latents processor plugins including conversion (interposer) / noise / scale
• Options for manipulation of pipeline sigmas, including built in expression parser for easy scaling
• patchmatch CPU based inpainting image processor implemented in Cython
• inpaint image processor (spandrel based MAT/LaMa)
• Segment Anything and YOLO mask generation / detection image processors via Ultralytics
• New image processors such as paste / crop-to-mask to aid with manual adetailer like editing tasks
• adetailer pipeline (adetailer on any image) / adetailer image processor for --post-processors usage
• image processor arch overhaul, use image processors at any pixel alignment
• SigLIP IP Adapter for SD3
• GPU memory management overhaul, mostly internal architecture
• Experimental XPU support
• Fix outdated timm library reliance in controlnet_aux
• Command line argument compression, but also even more arguments / parameterizations
• (Console UI) Cross platform OpenGL accelerated image viewer with pan/zoom, bounding box and point selection tools
• (Console UI) Countless UI improvements to Recipe / URI insertion UIs, a form for every possible URI argument
• (Console UI) A few new color schemes
• (Console UI) Many stability fixes
• Move to cross platform network installer instead of PyInstaller + WiX