Research map / Image generation / Fast diffusion sampling
Soft-Di[M]O: Improved one-step Image Discrete Model
Yuanzhi Zhu, Xi WANG, Stéphane Lathuilière, Vicky Kalogeiton
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 soft embeddings plus GAN refinement to Di[M]O | Di[M]O with MaskBit teacher · ImageNet-256 | FID: 2.89 -> 1.96 | Improved |
| train Soft-Di[M]O longer with slightly changed hyperparameters | Soft-Di[M]O MaskBit · ImageNet-256 | FID: 1.96 -> 1.56 | Improved |
| add soft embeddings and GAN to Di[M]O | Di[M]O with MaskBit teacher · ImageNet-256 | recall: 0.49 -> 0.55 | Improved |
| add soft embeddings and GAN to Di[M]O | Di[M]O with MaskBit teacher · ImageNet-256 | IS: 310.1 -> 281.4 | Hurt |
| add soft embeddings (GAN) to Di[M]O with MaskGit teacher | Di[M]O MaskGit · ImageNet-256 | FID: 6.91 -> 6.40 | Improved |
| add differentiable reward fine-tuning with Di[M]O loss as regulariser | Di[M]O MaskGen-L · Text-to-image GenEval overall | GenEval: 0.42 -> 0.51 | Improved |
| add differentiable reward fine-tuning | Di[M]O Meissonic · Text-to-image HPS V2.1 | HPS averaged: 28.59 -> 32.35 | Improved |
| add test-time embedding optimization (TTEO) | Soft-Di[M]O-MaskGen-L · Text-to-image GenEval overall | GenEval: 0.51 -> 0.63 | Improved |
| use soft embedding instead of Gumbel straight-through estimator (ablation) | Soft-Di[M]O · ImageNet-256 (5k FID) | FID | Improved |
| use GAN loss only without Di[M]O loss (ablation) | Soft-Di[M]O · ImageNet-256 and MaskGen-L | FID / diversity | Hurt |
| add Di[M]O loss as regulariser to the GAN loss (ablation) | Soft-Di[M]O · ImageNet-256 | FID | Improved |
| add random masking of soft embeddings for the discriminator, larger range (ablation) | Soft-Di[M]O GAN · ImageNet-256 | FID | Improved |
Closest papers by meaning
- RankE: End-to-End Post-Training for Discrete Text-to-Image Generation with Decoder Co-Evolution (NeurIPS 2026 · similarity 0.88)
- SoftVQ-VAE: Efficient 1-Dimensional Continuous Tokenizer (CVPR 2025 · similarity 0.87)
- Distilled Decoding 2: One-step Sampling of Image Auto-regressive Models with Conditional Score Distillation (NeurIPS 2025 · similarity 0.87)
- Unlocking Dataset Distillation with Diffusion Models (NeurIPS 2025 · similarity 0.87)
- Compositional Discrete Latent Code for High Fidelity, Productive Diffusion Models (NeurIPS 2025 · similarity 0.87)
- David and Goliath: Small One-step Model Beats Large Diffusion with Score Post-training (ICML 2025 · 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.
