Research map / Image generation / Autoregressive and latent generation
SpeeDiff: Scalable Pixel-Anchored End-to-End Latent Diffusion Model
Bingliang Zhang, Wenda Chu, Yizhuo Li, Linjie Yang, Yisong Yue, Katherine Bouman, Yang Song, Qiushan Guo
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 |
|---|---|---|---|
| backpropagate diffusion loss into VAE encoder without any pixel feedback (vanilla end-to-end) (ablation) | joint VAE + diffusion training · ImageNet 256x256, 80 epochs | gFID: 13.02 -> 33.95 (vs two-stage baseline) | Hurt |
| add Tweedie Pixel Reconstruction (TPR) loss to end-to-end training (ablation) | Vanilla E2E · ImageNet 256x256, 80 epochs | gFID: 33.95 -> 5.79 | Improved |
| add refined architecture (ViT-VAE and Refined-DiT) to E2E with TPR (ablation) | E2E w/ TPR · ImageNet 256x256, 80 epochs | gFID: 5.79 -> 3.66 | Improved |
| add REPA++ alignment to E2E (ablation) | SpeeDiff (w/o REPA++) · ImageNet 256x256, 80 epochs | gFID: 3.66 -> 1.69 | Improved |
| allow diffusion gradient to flow into VAE (end-to-end) versus stop-gradient, with all components (ablation) | SpeeDiff · ImageNet 256x256, 80 epochs | gFID: 3.47 -> 1.69 | Improved |
| add TPR loss under stop-gradient single-stage training (ablation) | Single-stage Baseline · ImageNet 256x256, 80 epochs | gFID: 14.21 -> 11.65 | Improved |
| use two-stage VAE-then-DiT vs single-stage with stop-gradient (ablation) | latent diffusion baseline · ImageNet 256x256, 80 epochs | gFID: 13.02 vs 14.21 | No clear effect |
| add Latent-REPA and Diff-REPA together (REPA++) instead of either alone (ablation) | SpeeDiff-XL · ImageNet 256x256, 80 epochs | gFID: 2.28 / 2.15 -> 1.69 | Improved |
| use DINOv3-L instead of DINOv2-L as the alignment VFM (ablation) | REPA++ · ImageNet 256x256, 80 epochs | gFID: 1.76 -> 1.69 | Improved |
| use 32 latent channels instead of 16, 64 or 128 (ablation) | ViT-VAE-XL + Refined-DiT-XL · ImageNet 256x256, 80 epochs | gFID: 2.82 (16), 2.27 (64), 3.26 (128) -> 1.69 (32) | Improved |
| increase latent channels (rFID improves but generation worsens at high counts) (ablation) | ViT-VAE-B + Refined-DiT-B · ImageNet 256x256, 80 epochs | gFID / rFID: rFID 0.84 -> 0.26 and gFID 6.20 -> 12.06 (32 -> 128 channels) | Mixed |
| train a new diffusion model on the frozen VAE learned by SpeeDiff (ablation) | SpeeDiff-XL · ImageNet 256x256, 80 epochs | gFID: 1.73 vs 1.69 | No clear effect |
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.
- REPA-E — REPA-E: Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers (ICCV 2025) · 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
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.91)
- PixelDiT2: Representation-Grounded Pixel Diffusion Transformers (NeurIPS 2026 · similarity 0.90)
- ScaleDiff: Higher-Resolution Image Synthesis via Efficient and Model-Agnostic Diffusion (NeurIPS 2025 · similarity 0.89)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.89)
- OSCAR: One-Step Diffusion Codec Across Multiple Bit-rates (NeurIPS 2025 · similarity 0.89)
- PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion (NeurIPS 2026 · similarity 0.88)
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.
