{"slug": "spectral-guided-diffusion-accelerating-inference-via-static-spectral-layer", "title": "Spectral-Guided Diffusion: Accelerating Inference via Static Spectral Layer Scheduling", "summary": "A new arXiv paper introduces the Spectral Concentration Ratio (SCR), an offline sensitivity proxy that assigns each residual branch a deterministic lifetime so pretrained diffusion models can skip recomputing frozen units during inference. At matched layer-step budgets, SCR combined with Frobenius magnitude preserved quality better than random, depth, norm, stable-rank, and Frobenius–stable-rank schedules on LLaDA-8B, DiT-XL/2, U-ViT-L, and SDXL, with the ranking holding down to ten denoising steps. The captured-graph system reached 2.8x–3.0x wall-clock speedup over eager inference, where padded graph execution alone accounted for 2.7x on LLaDA and eliminating inactive branch work raised it to 3.0x.", "body_md": "arXiv:2609.29505v1 Announce Type: new \nAbstract: Diffusion inference repeatedly evaluates the same large network. We ask whether pretrained weights alone can identify residual branches that need not be recomputed throughout the trajectory. Our \\textbf{Spectral Concentration Ratio (SCR)} measures leading-versus-tail singular-value energy. Combined with Frobenius magnitude, it yields an offline sensitivity proxy and a deterministic lifetime for each scheduled unit. A frozen unit reuses its cached residual-branch update while the current residual stream and all external conditioning continue to propagate. The method needs no router, calibration prompts, or input-dependent search. At matched layer-step budgets, SCR/Frobenius preserves quality better than random, depth, norm, stable-rank, and Frobenius--stable-rank schedules on LLaDA-8B, DiT-XL/2, U-ViT-L, and SDXL. Broader LLaDA tests cover retrieval, reasoning, code, summarization, and open-ended generation; matched-horizon controls retain the ranking down to ten denoising steps. The complete captured-graph system reaches $2.8\\times$--$3.0\\times$ wall-clock speedup over eager inference. This is a systems-level number: on LLaDA, padded graph execution already gives $2.7\\times$, while eliminating inactive branch work raises it to $3.0\\times$. The perturbation analysis motivates pre-norm attention and MLP components under explicit local assumptions; results on AdaLN, U-shaped, convolutional, and cross-attention blocks are empirical transfer, not certified guarantees.", "url": "https://wpnews.pro/news/spectral-guided-diffusion-accelerating-inference-via-static-spectral-layer", "canonical_source": "https://www.machinebrief.com/news/spectral-guided-diffusion-accelerating-inference-via-static-nvvx", "published_at": "2026-09-25 04:00:00+00:00", "updated_at": "2026-09-25 05:01:02.521816+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "generative-ai", "ai-research", "ai-infrastructure"], "entities": ["Spectral Concentration Ratio", "LLaDA-8B", "DiT-XL/2", "U-ViT-L", "SDXL", "arXiv"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/spectral-guided-diffusion-accelerating-inference-via-static-spectral-layer", "markdown": "https://wpnews.pro/news/spectral-guided-diffusion-accelerating-inference-via-static-spectral-layer.md", "text": "https://wpnews.pro/news/spectral-guided-diffusion-accelerating-inference-via-static-spectral-layer.txt", "jsonld": "https://wpnews.pro/news/spectral-guided-diffusion-accelerating-inference-via-static-spectral-layer.jsonld"}}