Research map / Image generation / Fast diffusion sampling
Analog Bits: Generating Discrete Data using Diffusion Models with Self-Conditioning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 Self-Conditioning at sampling and training (ablation) | Bit Diffusion / DDPM, DPM-Solver 20 NFE (x0 prediction) · ImageNet 64x64 continuous | FID: 12.13 -> 6.94 | Helped |
| add Self-Conditioning (eps prediction) (ablation) | continuous diffusion · ImageNet 64x64 continuous, DPM-Solver 20 NFE | FID: 6.10 -> 4.24 | Helped |
| add Self-Conditioning on binary encodings (ablation) | Bit Diffusion · CIFAR-10 | FID | Helped |
| add asymmetric time intervals (time difference 0 -> 7/8) at sampling (ablation) | Bit Diffusion captioning, 10 steps · MS-COCO captioning | BLEU-4: 17.6 -> 34.6 | Helped |
| add asymmetric time intervals (ablation) | Bit Diffusion captioning, 5 steps · MS-COCO captioning | BLEU-4: 17.1 -> 31.6 | Helped |
| add non-zero time difference at 400 steps (ablation) | Bit Diffusion · ImageNet 64x64 | FID | No clear effect |
| add non-zero time difference at 100 steps (ablation) | Bit Diffusion · ImageNet 64x64 | FID | Helped |
| use uint8 (rand) categorical encoding instead of uint8 (ablation) | Bit Diffusion · ImageNet 64x64 class-conditional | FID: 4.84 -> 8.76 | Hurt |
| replace base-2 uint8 with one-hot encoding (ablation) | Bit Diffusion l2 loss · CIFAR-10 unconditional | FID: 3.48 -> 46.32 | Hurt |
| replace l2 loss with logistic loss (ablation) | Bit Diffusion uint8 (rand) · CIFAR-10 unconditional | FID: 6.93 -> 49.29 | Hurt |
| replace l2 loss with logistic loss (ablation) | Bit Diffusion uint8 / gray code · CIFAR-10 unconditional | FID: 3.48 -> 3.53; 3.86 -> 3.71 | No clear effect |
| use self-guidance sampling instead of default Self-Conditioning sampling (ablation) | Bit Diffusion · ImageNet 64x64 class-conditional | FID: uint8 (rand) 8.76 -> 7.87; uint8 4.84 -> 5.15 | Mixed |
Closest papers by meaning
- BitsFusion: 1.99 bits Weight Quantization of Diffusion Model (NeurIPS 2024 · similarity 0.90)
- BiDM: Pushing the Limit of Quantization for Diffusion Models (NeurIPS 2024 · similarity 0.88)
- Compressed Image Generation with Denoising Diffusion Codebook Models (ICML 2025 · similarity 0.88)
- Simplified and Generalized Masked Diffusion for Discrete Data (NeurIPS 2024 · similarity 0.87)
- DCTdiff: Intriguing Properties of Image Generative Modeling in the DCT Space (ICML 2025 · similarity 0.87)
- Autoregressive Image Generation without Vector Quantization (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.