DiffusionBear v0.4.1 beta

🌟 DiffusionBear v0.4.1 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.

🍎 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 either run:

xattr -cr /Applications/DiffusionBear.app

or use System Settings → Privacy & Security → Open Anyway.

🌟 Release Highlights: SDXL Performance Recovery & Unified Memory Optimization

DiffusionBear v0.4.1 resolves SDXL memory thrashing on Apple Silicon Macs, elevates generation throughput, and introduces our complete empirical 187-generation multi-model benchmark suite.

• SDXL Memory Thrashing Root-Cause Resolution (+51.9% Suite Acceleration):

  • The Issue: Persistent daemon caching left up to 7.8 GB of stale Metal allocations resident between runs. When clamped against an artificial 6.5 GB limit, subsequent runs suffered from severe Metal buffer evictions and paging stalls.
  • The Fix: Systematic cache purging (gc.collect() and mx.clear_cache()) inside execution finally: blocks and daemon loops, combined with removing artificial clamps in favour of dynamic system working-set allocation.
  • Results: Global SDXL suite generation time plunged from 4,569.5s (1h 16m 09s) down to 2,195.9s (36m 35s): +51.9% faster across the board, saving over 39 minutes of compute.
    • Juggernaut XL Lightning: 4.25x faster (+76.5%), down from 423.2s to 99.6s (strictly 9.15s–9.30s per image).
    • RealVisXL V5.0: 3.63x faster (+72.5%), down from 1,709.9s to 470.8s.
    • Juggernaut XI: 2.15x faster (+53.4%), down from 988.3s to 460.8s.
    • Pony: +19.6% faster, down from 1,448.0s to 1,164.7s.

• Bit-Identical Visual Fidelity Recovery (∞ dB PSNR):

  • Distilled SDXL models (4-step Lightning, 8-step Hyper-SD) require trailing timesteps down to $t=0$. DiffusionBear now automatically promotes them to euler_trailing sampling.
  • Verification: Juggernaut XL Lightning achieved ∞ dB PSNR (0.00 / 255 mean pixel difference) against our golden baseline — restoring 100% bit-for-bit mathematical visual parity.

• High-Performance TAESD VAE Decoding:

  • Integrated standalone taesdxl autoencoder for ultra-lightweight, sub-second latent decode (~0.45s), removing VAE decoding bottlenecks.

• Complete 187-Run Empirical Benchmark Arena:

  • Dual-benchmark architecture preserving legacy data while introducing all 17 modern configurations across 4 engine families (FLUX, Krea, DiT, SDXL) with an interactive side-by-side comparison slider on both site/ and docs/.

• Custom Model Dimensions & Reactive Dimension Tags:

  • The Parameters tab now exposes and displays exact default dimensions (width and height) for each model. Users can customize and persist model-specific defaults.
  • Active custom dimensions dynamically reflect in the Generate tab with a clear "Custom" tag indicator in the model picker and size selector.

• High-Contrast Dark Mode Form Affordance:

  • Explicit color-scheme: dark; configuration and high-contrast WebKit stepper filters restore crisp visibility to up/down spinners (input[type="number"]), select arrows, and navigation chevrons.

📦 Checksums

shasum -a 256 DiffusionBear-0.4.1-arm64.zip
097ccacd46cd3b7707389dadcbd865ea5c4a7b3c5b71f58d5c409d60880ae08c

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