Research map / Image generation / Autoregressive and latent generation
Hi-DiT: Hybrid Latent-Pixel Diffusion Transformer for Image Generation
Yedong Shen, Yehao Li, Yingwei Pan, Yanyong Zhang, Ting Yao
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 parallel latent branch to pixel-stream baseline (ablation) | Hi-DiT · ImageNet 256x256 (80-epoch ablation) | FID: 3.38 -> 1.74 | Improved |
| replace MLP pixel predictor with hierarchical sub-pixel HF Pixel Predictor (ablation) | Hi-DiT · ImageNet 256x256 (80-epoch ablation) | FID: 1.74 -> 1.65 | Improved |
| set time-gating threshold tau to 0.3 instead of 0.7 (ablation) | Hi-DiT · ImageNet 256x256 (80-epoch ablation) | FID: 1.73 -> 1.65 | Improved |
| set time-gating threshold tau to 0.3 instead of 0.1 (ablation) | Hi-DiT · ImageNet 256x256 (80-epoch ablation) | FID: 1.75 -> 1.65 | Improved |
| add latent+pixel hybrid with time-gated injection to SiT latent diffusion | SiT-XL · ImageNet 256x256 (w/ CFG) | FID: 2.06 -> 1.06 | Improved |
| hybrid latent-pixel design without CFG | SiT-XL · ImageNet 256x256 (w/o CFG) | FID: 8.61 -> 1.66 | Improved |
| add hybrid latent-pixel design to U-ViT-S/2 for text-to-image | U-ViT-S/2 · MS-COCO 256x256 text-to-image | FID: 5.48 -> 5.15 | Improved |
| raise CFG scale from 2.4 to 3.0 (ablation) | Hi-DiT · ImageNet 256x256 (w/ CFG) | FID / IS: FID 1.06 -> 1.10, IS 296.6 -> 319.4 | Mixed |
| add pixel stream and pixel predictor on top of latent backbone | SiT / VA-VAE · ImageNet 256x256 cost | inference time per image (s): 1.52 -> 1.67 (SiT); peak mem 32.90G -> 35.52G | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PixelDiT — PixelDiT: Pixel Diffusion Transformers for Image Generation (CVPR 2026) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- PixelDiT2: Representation-Grounded Pixel Diffusion Transformers (NeurIPS 2026 · similarity 0.91)
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.90)
- Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.90)
- Scalable High-Resolution Pixel-Space Image Synthesis with Hourglass Diffusion Transformers (ICML 2024 · similarity 0.90)
- Dynamic Diffusion Transformer (ICLR 2025 · similarity 0.89)
- Diffusion Models Without Attention (CVPR 2024 · similarity 0.89)
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.
