Research map / Image generation / Autoregressive and latent generation
Pyramid Patchification Flow for Visual Generation
Hui Li, Baoyou Chen, Li jiaye, Jingdong Wang, Siyu Zhu
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 |
|---|---|---|---|
| use pyramidal patchification (larger patches at high noise), two stages | SiT-B/2 trained from scratch · ImageNet 256 | FID-50k: 4.46 -> 3.83 at 62.0 testing FLOPs (%) | Improved |
| use three-stage pyramidal patchification | SiT-B/2 trained from scratch · ImageNet 256 | FID-50k: 4.46 -> 4.43 | No clear effect |
| use pyramid representations instead of full-resolution latents (ablation) | PPFlow · ImageNet 256, 1M steps from scratch | FID-50k: 15.68 -> 164.48 | Hurt |
| add renoising trick to pyramid representations (ablation) | Pyramidal Flow variant · ImageNet 256, 1M steps | FID-50k: 164.48 -> 27.69 | Improved |
| train each patch size only on its own timestep range instead of all timesteps (Lumina-Video style) (ablation) | PPFlow · ImageNet 256, 1M steps | FID-50k: 18.77 -> 15.68 | Improved |
| add learned patch-level (stage) embedding (ablation) | PPF-B-2 · ImageNet 256, 1M steps from SiT-B/2 | FID-50k: 4.44 -> 4.30 | Improved |
| add stage-wise CFG schedule (ablation) | PPF-B-2 · ImageNet 256, 1M steps from SiT-B/2 | FID-50k: 4.30 -> 4.22 | Improved |
| increase number of stages from 2 to 5 with equal 1M-step training (ablation) | PPF-B · ImageNet 256, from SiT-B/2 | FID-50k: 4.22 -> 5.30 | Hurt |
| train five-stage PPFlow longer (ablation) | PPF-B-5 · ImageNet 256, from SiT-B/2 | FID-50k: 5.30 -> 4.51 | Improved |
| move first stage boundary earlier (more full-resolution steps) (ablation) | PPF-B-3 · ImageNet 256 | FID-50k / testing FLOPs: FID 4.57 -> 4.21; FLOPs 49.1 -> 68.3 | Mixed |
| apply PPFlow to FLUX.1-dev text-to-image | FLUX.1-ft · GenEval, DPG Bench, T2I-CompBench at 512 to 2048 | GenEval: 0.68 -> 0.68 at 62.2 FLOPs (%) (512) | No clear effect |
| use PPFlow instead of FlexiDiT at about 63% FLOPs | DiT-XL/2 · ImageNet 256 | FID-50k: 2.25 -> 2.15 | Improved |
Closest papers by meaning
- Pyramidal Flow Matching for Efficient Video Generative Modeling (ICLR 2025 · similarity 0.89)
- Improving Progressive Generation with Decomposable Flow Matching (NeurIPS 2025 · similarity 0.89)
- Hierarchical Patch Diffusion Models for High-Resolution Video Generation (CVPR 2024 · similarity 0.88)
- FlexiDiT: Your Diffusion Transformer Can Easily Generate High-Quality Samples with Less Compute (CVPR 2025 · similarity 0.87)
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.87)
- Patched Denoising Diffusion Models For High-Resolution Image Synthesis (ICLR 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.
