This is a long overdue release. I had implemented a bunch of things earlier this year and then subsequently forgot to release them after they had been tested out by people, so this update is releasing all that to stable.

Part of why I hadn't released these changes is because I have been privately working on a Rust rewrite of chaiNNer's backend. I've been working on and off on that since around March. I will discuss that more soon, so stay tuned for updates and announcements on that. If you don't know what that entails, basically that just means everything will be significantly faster, and some things (like dependency management) a bit less jank.

This is the last release (minus any patches for necessary bug fixes) with the Python backend. It is also the last release that will be marked Alpha, and going forward I will use true semantic versioning starting from v27.0.0.

Major Changes

• TensorRT upscaling and engine conversion, with BF16 support. Requires an RTX 20xx / GTX 16xx GPU or newer. (#3247, #3265, #3254, #3255)
• Language switching: the interface is now available in English, German, Spanish, French, Hindi, Italian, Japanese, Korean, Dutch, Polish, Portuguese (Brazil), Russian, Turkish and Chinese (Simplified). Change it in Settings → Appearance (#3171, #3180, #3341)
  • Node names, descriptions and inputs, the app menu, some dialogs and messages, and type descriptions are still in English
  • The translations are machine-made and machine-reviewed. Corrections from native speakers are welcome.

New nodes and node features

• New Math Expression node: write a formula with variable inputs (#3187, #3191)
• New Is Grayscale node: checks whether an image is grayscale (#3222)
• Regex Find has a new "Found" output, so a chain no longer fails when nothing matches (#3140)
• Channel nodes are renamed to consistent "Split" and "Combine" names (#3135)

Minor Changes

• Upscale nodes show their own progress bar while processing tiles, for PyTorch, ONNX, NCNN and TensorRT (#3252)
• Lower CPU usage from progress updates during long runs (#3250)
• PyTorch auto split takes the model's size requirements into account (#3167)
• Installation progress shows on the taskbar (#3169)
• Installing no longer hangs silently when the disk runs out of space (#3131)
• Python 3.8 and 3.9 are no longer supported, because both have reached end of life. If you use your own system Python, it must be 3.10 or newer (#3207)
• The backend now writes independent log files, which helps with bug reports. If submitting a bug, please send all three log files: main.log, host.log and worker.log (#3137)
• Updated dependencies

Bug Fixes

• Fixed VRAM usage growing every time the PyTorch inpaint node ran (#3194)
• Fixed cached node results never being freed (#3199)

Full Changelog: v0.26.0...v0.26.0