Research map / Image generation / Autoregressive and latent generation
DeCo: Frequency-Decoupled Pixel Diffusion for End-to-End Image Generation
Zehong Ma, Longhui Wei, Shuai Wang, Shiliang Zhang, Qi Tian
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 lightweight pixel decoder for high-frequency detail, DiT handles low-frequency semantics (ablation) | pixel diffusion baseline (DiT-L) · ImageNet 256x256, 200k iterations | FID: 61.10 -> 34.12 | Improved |
| add frequency-aware flow-matching loss (ablation) | DeCo · ImageNet 256x256, 200k iterations | FID: 34.12 -> 31.35 | Improved |
| add REPA alignment loss (ablation) | DeCo · ImageNet 256x256, 200k iterations | FID: 67.55 -> 31.35 | Improved |
| use AdaLN interaction instead of upsample-and-add (ablation) | DeCo · ImageNet 256x256, 200k iterations | FID: 36.02 -> 31.35 | Improved |
| use pixel decoder patch size 1 instead of 16 (ablation) | DeCo · ImageNet 256x256, 200k iterations | FID: 55.59 -> 31.35 | Improved |
| use pixel decoder patch size 1 instead of 4 (ablation) | DeCo · ImageNet 256x256, 200k iterations | FID: 34.39 -> 31.35 | Improved |
| increase pixel decoder hidden size from 32 to 64 (ablation) | DeCo · ImageNet 256x256, 200k iterations | FID: 34.12 -> 35.88 | No clear effect |
| increase pixel decoder depth from 3 to 6 (ablation) | DeCo · ImageNet 256x256, 200k iterations | FID: 34.12 -> 35.46 | Hurt |
| reduce pixel decoder depth from 3 to 1 (ablation) | DeCo · ImageNet 256x256, 200k iterations | FID: 34.12 -> 37.10 | Hurt |
| set JPEG quality to 100 (equal frequency weights) in frequency-aware loss (ablation) | DeCo · ImageNet 256x256, 200k iterations | FID: 31.35 -> 33.84 | Hurt |
| reduce JPEG quality to 50 (ablation) | DeCo frequency-aware loss · ImageNet 256x256, 200k iterations | FID: 31.35 -> 31.54 | No clear effect |
| change frequency loss weight from 1 to 0.5 or 2 (ablation) | DeCo · ImageNet 256x256, 200k iterations | FID: 31.35 -> 33.54 (0.5); 32.97 (2) | Hurt |
Closest papers by meaning
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.90)
- PixelDiT2: Representation-Grounded Pixel Diffusion Transformers (NeurIPS 2026 · similarity 0.90)
- PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion (NeurIPS 2026 · similarity 0.90)
- FAM Diffusion: Frequency and Attention Modulation for High-Resolution Image Generation with Stable Diffusion (CVPR 2025 · similarity 0.90)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.89)
- PixelART: Image-to-Layer Decomposition without Latents or Text-to-Image Pretraining (NeurIPS 2026 · 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.
