Research map / Image generation / Fast diffusion sampling

DC-DiT: Adaptive Compute and Elastic Inference for Visual Generation via Dynamic Chunking

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Training-free diffusion acceleration · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DC-DiT, a diffusion transformer that replaces fixed patchify with a learned encoder-router-decoder that dynamically chunks latent tokens, enabling elastic inference via tail dropping and Lite-CFG, and reports better FID-FLOPs tradeoffs on ImageNet and an upcycled Z-Image.

Problem

Diffusion Transformers use static patchify tokenization, giving the same token budget to every region and denoising timestep regardless of content, which wastes compute.

Approach

A learned router predicts per-token boundary probabilities from local spatial unpredictability; only boundary tokens go through the DiT backbone, trained end-to-end with a straight-through estimator, while tail dropping of low-confidence boundaries with multi-budget training yields elastic inference from a single checkpoint.

Key results

• DC-DiT-XL at 512x512 with Lite-CFG reduces compute by 36.8% (522.50 to 330.00 TFLOPs/img) with FID 12.50 vs 11.04 for DiT-XL/2.
• Under Lite-CFG at 256x256, DC-DiT-B improves FID from 29.72 to 18.49 (-37.8%) while cutting TFLOPs/img by 19.8%.
• Upcycled DC-Z-Image at 60% tail drop reduces cost from 5603 (Z-Image) to 2570 TFLOPs/img while keeping DPG-Bench at 86.32 vs 85.97 for Z-Image.

Surprising findings

• Baseline DiT-S/2 and DiT-L/2 at 512px repeatedly diverged during training, whereas DC-DiT-S and DC-DiT-L trained stably.
• Without tail dropping, DC-DiT-XL at 256px is slightly worse than DiT-XL/2 (FID 19.73 vs 19.47) despite higher FLOPs, so gains are concentrated in lower-compute regimes and Lite-CFG settings.

Limitations

• Gains on XL scale are mostly competitive rather than superior to the DiT baseline at full compute.
• Compute-matched training requires extending DC-DiT training to 400K-495K steps, so the comparison depends on the FLOPs accounting protocol.

Experimental setup

Datasets: ImageNet Recap-DataComp-1B prompts

Benchmarks: ImageNet 256x256 ImageNet 512x512 DPG-Bench

Metrics: FID-50K TFLOPs/img DPG-Bench score

Baselines: DiT-S/2 DiT-B/2 DiT-L/2 DiT-XL/2 Z-Image DyDiT TeaCache

Builds on: DiT H-Net DyDiT TeaCache FlashAttention

Closest papers by meaning

• Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.89)
• When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.88)
• UDT: Reconciling U-Nets and Diffusion Transformers with Data-Adaptive Token Reduction (NeurIPS 2026 · similarity 0.87)
• SparseDiT: Token Sparsification for Efficient Diffusion Transformer (NeurIPS 2025 · similarity 0.87)
• Dynamic Diffusion Transformer (ICLR 2025 · similarity 0.87)
• Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.87)