SOTA Architecture Benchmark: JIT Compilation & Apple Silicon Metal Optimization

🔬 1. DeepCache Empirical Analysis: Distillation Limits & 12-Step Resolution

DeepCache exploits temporal redundancy in UNet diffusion architectures by caching high-level feature maps from early downsampling blocks and reusing them across consecutive denoising steps. While highly effective on unconstrained, high-step models (25–50 steps), our empirical benchmarks demonstrate severe trade-offs when applied to distilled trajectories:

• 4-Step Distilled Trajectories (SDXL Lightning): Skipping UNet forward passes in a 4-step schedule leads to catastrophic structural degradation. High-frequency textures (such as animal fur, micro-gears, and fine water ripples) lose definition, dropping PSNR below 23 dB. Increasing step count to compensate leads to overcooked contrast. Consequently, DeepCache remains strictly disabled by default in DiffusionBear.
• 12-Step Standard Trajectories: On 12 steps, DeepCache achieves a 1.47x speedup with manageable fidelity loss (PSNR ~25.5 dB). It is provided as an advanced, optional toggle for draft exploration only.

🏆 2. SOTA Architecture Evaluation: DiT vs UNet vs Qwen on Apple Silicon

The shift from traditional UNets to Diffusion Transformers (DiTs) brings unprecedented spatial composition and prompt adherence to local Mac inference:

• FLUX.3-klein 5.6B: Established as our new primary flagship. Operating at 13.3 s/step with native flow_match_euler_discrete scheduling, it delivers surgical typography and complex multi-object spatial coherence in 53.3s.
• Krea 2 Turbo 13B: Shaved over 75.7 seconds in v0.4.1 via our custom 2.26 GB Q4-quantized text encoder persistence, dropping from 250.8s to 175.1s with zero bf16 load overhead.
• Z-Image Turbo 6B: Demonstrated our largest single optimization gain: dropping from 170.4s to 120.47s (+26.5% speedup) through fused JIT step kernels.

⚙️ 3. Apple Silicon Metal Runtime & SDXL Engineering Findings

Achieving uninterrupted 11.8-hour benchmark runs without a single OOM or memory leak required deep runtime hardening:

• JIT Compilation with @mx.compile: Fusing Euler and Flow-Matching step loops directly into Metal compute kernels eliminated Python-GPU roundtrip synchronization stalls, yielding immediate, 100% lossless speed gains across all diffusion models.
• Unified Memory Budgeting (10.88 GB Bound): Capping the wired memory budget at 68% of unified RAM guarantees that the Metal driver and macOS window server never compete for swap space, maintaining responsive system performance throughout sustained generation batches.
• SDXL Runtime Architecture: Running SDXL UNet in 4-bit quantization (quantize_unet=4) is 5x faster than fp16 on M1, whereas 8-bit quantization is 7x slower and explicitly prohibited. Furthermore, IPC pipe handling in sdxl_engine.py and daemon lifecycle managers in generator.py were hardened to cleanly reap background workers and prevent dangling child processes.