Research map / Image generation / Autoregressive and latent generation
Training-free Mixed-Resolution Latent Upsampling for Spatially Accelerated Diffusion Transformers
Wongi Jeong, Kyungryeol Lee, Hoigi Seo, Se Young Chun
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add NT-Matching with optimized JSD (0.026 vs 0.040) (ablation) | RALU · FLUX 7x | ImageReward: 0.966 -> 0.999 | Improved |
| add NT-Matching with optimized JSD (0.026 vs 0.030) (ablation) | RALU · FLUX 7x | T2I-CompBench: 0.571 -> 0.633 | Improved |
| use VAE decode + Canny edge detection instead of latent gradient kernel (ablation) | RALU · FLUX 5x | ImageReward: 0.996 -> 1.022 | Improved |
| use adaptive upsampling ratio instead of fixed ratio r=0.2 (ablation) | RALU · FLUX 5x | ImageReward: 1.015 vs 1.022 | No clear effect |
| increase early upsampling ratio from low values toward 0.3 (ablation) | RALU · FLUX (ImageReward and CLIP-IQA vs TFLOPs) | image quality | Mixed |
| upsample edge regions early instead of upsampling all latents late (Bottleneck Sampling) | spatial acceleration · FLUX 5x | NIQE: 9.16 -> 6.43 | Improved |
| combine RALU with TaylorSeer (W=3) | RALU · FLUX 5x | speedup: 5.53x -> 7.28x | Improved |
| combine RALU with Delta-DiT | RALU · FLUX 5x | ImageReward: 1.022 -> 0.827 | Hurt |
| combine RALU with ToCa | RALU · FLUX 5x | CLIP-IQA: 0.700 -> 0.522 | Hurt |
| apply RALU to timestep-distilled FLUX.1-schnell | FLUX.1-schnell · FLUX.1-schnell 4 steps | speedup: 11.83x -> 15.91x | Improved |
| apply RALU to timestep-distilled FLUX.1-schnell | FLUX.1-schnell · FLUX.1-schnell 4 steps | ImageReward: 1.055 -> 0.992 | Hurt |
| apply RALU to SD3.5L-Turbo | SD3.5L-Turbo · SD3.5L-Turbo 4 steps | ImageReward: 0.840 -> 0.789 | Hurt |
Closest papers by meaning
- FasterDiT: Towards Faster Diffusion Transformers Training without Architecture Modification (NeurIPS 2024 · similarity 0.90)
- Accelerating Diffusion Transformer via Increment-Calibrated Caching with Channel-Aware Singular Value Decomposition (CVPR 2025 · similarity 0.90)
- SparseDiT: Token Sparsification for Efficient Diffusion Transformer (NeurIPS 2025 · similarity 0.88)
- U-DiTs: Downsample Tokens in U-Shaped Diffusion Transformers (NeurIPS 2024 · similarity 0.87)
- BlockDance: Reuse Structurally Similar Spatio-Temporal Features to Accelerate Diffusion Transformers (CVPR 2025 · similarity 0.87)
- On Statistical Rates and Provably Efficient Criteria of Latent Diffusion Transformers (DiTs) (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
