Repository navigation
DiffusionBear 0.3.8 (beta) — stability pass across every engine
A stability release. Every engine in the registry was re-verified by live generation on a MacBook Pro (Apple M1, 16 GB).
Fixed
shapeless=Truecompilation removed from FLUX.2 (txt2img + edit), Krea 2 and Z-Image. Shapeless compilation cannot infer output shapes for aSplitwhose axis length is only known at runtime — Krea 2 died at denoise step 3/8 withSplit cannot infer output shapes. Compilation is now shape-specialised, keeping the graph fusion without the crash class. Costs ~6 s on a 512×768 FLUX.2 4-step run.- FLUX.2 reference-image editing failed at denoise step 3/4 with
[eval] Attempting to eval an array during function transformations. Root cause was our rope memo cache, which forced an evaluation while the graph was being traced. The cache now detects that case and falls through to the uncached call, so the edit path compiles again rather than being disabled to hide the symptom. - Qwen-Image 2.1 single-file text-encoder layouts returned
Nonefrom the weight selector, producing an opaqueAttributeError. All three layouts are handled now. - Installed models were being re-downloaded: the resolver missed the Hugging Face cache and treated interrupted
.incomplete/.parttransfers as finished. - FLUX.2-klein 4B edit could fall back to the unquantised 22 GB upstream repo instead of the 4-bit build.
Also
LSMinimumSystemVersionis now derived from the payload binaries (macOS 26.2), so the plist can no longer understate the requirement by 11 macOS majors.- The vendored mflux is reconciled with the tested development venv, so what CI runs is what ships.
- Dead diffusers engine, a duplicate monkey-patch and the temporary
/halloweenroutes removed. - Regression tests added for the rope cache and the Qwen selector layouts.
Install
Right-click the app → Open → Open. It is ad-hoc signed, not notarised, so Gatekeeper blocks a plain double-click the first time.
SHA-256 of DiffusionBear-0.3.8-arm64.zip:
4d63e2ab034745479175a92f4e35e251e4fd56599b0f1d018eed9b6482a3bba5
Apple Silicon only, 16 GB unified memory recommended. Bugs: https://github.com/OuincheWinch/DiffusionBear/issues