Skip to content

DiffusionBear v0.4.2 beta

Latest
Latest

Choose a tag to compare

View all tags
@OuincheWinch OuincheWinch released this 26 Oct 11:22
v0.4.2
d6ad7b9

🌟 DiffusionBear v0.4.2 beta

DiffusionBear v0.4.2 beta

Apple Silicon only (M1–M4). Intel Macs are not supported. Requires macOS 26.2 or later (derived from the shipped binaries' minimum-OS floor). 16 GB unified memory recommended.

⚠️ Post-Mortem & Incident Report: v0.4.2 Packaging & Dependency Bloat (Full Transparency)

We believe in complete engineering transparency with our users. Here is what happened with the v0.4.2 release packaging:

1. What Happened (The Initial Bug)

  • The initial v0.4.2 build published on 2026-10-26 (465.6 MB, SHA-256 cde8d5..., ~20 early downloads) worked flawlessly for FLUX.2, SDXL, Krea 2, and Z-Image Turbo.
  • However, when attempting to generate with the newly introduced FIBO 2.0 Edit Turbo, the app immediately failed with Generation failed: ModuleNotFoundError: No module named 'mlx_vlm'.
  • Root Cause: During local development, FIBO was tested against a local ./venv-fibo/ virtualenv. To keep the app bundle lean, this folder was not copied into the .app. Instead, FIBO runs on the main bundled runtime Resources/venv. However, mlx-vlm was omitted from packaging/build_runtime_venv.sh, so users on a clean install lacked the module.

2. The Second Mistake: The 139 MB Dependency Bloat

  • When adding mlx-vlm>=0.7.10 to the build script, running a standard pip install pulled the package's default setup.py dependencies.
  • mlx-vlm is designed as a standalone CLI/training/benchmark toolkit, so it declared heavy external dependencies: opencv-python (cv2: 119 MB), datasets -> pyarrow (122 MB) + pandas (41 MB), scipy (80 MB), and matplotlib (20 MB) — totaling 382 MB of unused code.
  • This produced an inflated temporary package of 604.6 MB (SHA-256 9a22c6...).
  • Why this was unacceptable: DiffusionBear uses native Apple Silicon MLX and Pillow/PIL for all image manipulation and compositing. None of cv2, pyarrow, pandas, scipy, or matplotlib are ever imported or used by our engine. Shipping 382 MB of dead weight was caught immediately by the team.

3. The Resolution: Clean 456 MB Release (--no-deps)

  • packaging/build_runtime_venv.sh was updated to install mlx-vlm with --no-deps: 
    pip install --quiet --no-deps "mlx-vlm>=0.7.10"
    All real runtime requirements (mlx, torch, Pillow, numpy, transformers, huggingface-hub) were already present in the bundle.
  • The bloated packages were completely excised.
  • The clean, verified package is now 456.0 MB (even lighter than the initial build):
    • SHA-256: 62794333b85a164112d065f8cdf24a6310964a22953acb47aa1e46a29440220e

4. Audit Table of All 3 Iterations

Build Iteration Size SHA-256 Checksum Status
v0.4.2 Clean Release (Current) 456.0 MB 62794333b85a164112d065f8cdf24a6310964a22953acb47aa1e46a29440220e ✅ Recommended for all users
v0.4.2 Bloated Build (Temporary) 604.6 MB 9a22c6bd35b1d49fd858830dfdd4ffaf096f83f977e9d52c856358fcffc7ccc4 ⚠️ Functional but carries 382 MB bloat
v0.4.2 Initial Release (~20 downloads) 465.6 MB cde8d562fbbfaafa91de1acf8e80e36de2d8330b2824498569ec1c09b2a73308 ❌ Missing mlx_vlm (FIBO generation fails)

⚡ If you downloaded the Initial Build (465 MB) and don't want to re-download:

🍎 First Launch on macOS (Gatekeeper)

DiffusionBear is ad-hoc signed, so Gatekeeper blocks the first double-click of a downloaded copy — one-time, expected. Move DiffusionBear.app to /Applications, then run:

xattr -cr /Applications/DiffusionBear.app

or System Settings → Privacy & Security → Open Anyway.

🚀 Highlights: Full Support for Bria FIBO 2.0 Edit Turbo (MLX 4-bit)

DiffusionBear v0.4.2 introduces native Apple Silicon support for Bria AI's FIBO 2.0 Edit Turbo via our first-ever 4-bit affine quantized port (Ouinche/FIBO2-Edit-Turbo-MLX-4bit).

  • Full Support for FIBO 2.0 Edit Turbo MLX 4-bit:
    • First-ever 4-bit quantized port of Bria AI's state-of-the-art instruction editing model running natively on Apple Silicon via MLX.
    • SOTA text-guided edits, multi-reference scene synthesis, and precision inpainting in only 4 denoising steps (~85s on baseline M1, ~25-30s on M4/M5).
    • Model weights on Hugging Face: Ouinche/FIBO2-Edit-Turbo-MLX-4bit (quantized from briaai/Fibo-2.0-edit-turbo).
  • Text-Guided Single Reference Edit:
    • High-fidelity single reference editing with natural language prompts (e.g. "the jacket color is now red").
    • Grounded by the 4-bit Qwen3-VL vision-language encoder preserving facial identity, posture, and ambient illumination.
  • Studio Masked Edit (Precision Inpainting with 0 Variance Outside Mask):
    • Interactive brush painting directly on the reference image backdrop in DiffusionBear Studio.
    • Zero Variance Guarantee: Unmasked areas are strictly preserved from the source reference image with deterministic 1.5px Gaussian edge feathering and alpha compositing.
    • Conforms to the official FIBO 2.0 inpainting protocol (masked region mapped to neutral mid-grey RGB(128, 128, 128) before conditioning).
  • Multiple References (Multimodal In-Context Editing):
    • Combine up to 4 reference images (e.g., character portrait + specific dress or object).
    • Seamlessly transfers garments, objects, and textures into new coherent photographic compositions.
  • 3-Stage Sequential Memory Management (< 6.8 GB Peak VRAM):
    • Lazy component lifecycle: loads 4-bit Qwen3-VL text encoder → extracts visual conditioning → evicts from RAM; loads 4-bit DiT transformer → denoises 4 steps → evicts from RAM; loads fp16 FLUX.2 VAE → decodes image.
    • Peak resident VRAM remains strictly below 6.8 GB on 16 GB Apple Silicon Macs with zero SSD swap.
  • Bundled Standalone Runtime (mlx_vlm included, zero bloat):
    • Full self-contained Python runtime bundling mlx_vlm without unused bloat inside DiffusionBear.app/Contents/Resources/venv.
  • Interactive Showcase & Telemetry:

📦 Checksums & Assets Inventory

Asset Name Description Size SHA-256 Checksum
DiffusionBear-0.4.2-arm64.zip (Clean) Clean standalone application (bundles mlx_vlm, no bloat) 456.0 MB 62794333b85a164112d065f8cdf24a6310964a22953acb47aa1e46a29440220e
DiffusionBear-0.4.2-fibo-offline-patch.zip Offline drop-in patch for early downloaders 4.1 MB 79071450c9daa3fabcca4f3adb38a8f37ee94f091ed340a6776de9210a853c88
DiffusionBear-0.4.2-fibo-hotfix.command 1-click terminal hotfix script 1.5 KB 23414b1e3422f3224c8d5e919d6a445ad2a084771c7137a2df00e8154f97a035

Model weights are not bundled; engines download on first use into ~/Library/Application Support/DiffusionBear.

Assets 5
Loading