Research map / Image generation / Fast diffusion sampling
Neon: Negative Extrapolation From Self-Training Improves Image Generation
Sina Alemohammad, Zhangyang Wang, Richard Baraniuk
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 |
|---|---|---|---|
| apply Neon negative extrapolation to self-trained EDM-VP | EDM-VP · CIFAR-10 conditional | FID: 1.78 -> 1.38 | Improved |
| apply Neon to EDM-VP | EDM-VP · FFHQ-64 | FID: 2.39 -> 1.12 | Improved |
| apply Neon to flow matching | Flow matching · CIFAR-10 | FID: 3.5 -> 2.32 | Improved |
| apply Neon to xAR-L | xAR-L · ImageNet-256 | FID: 1.28 -> 1.02 | Improved |
| apply Neon to VAR-d16 | VAR-d16 · ImageNet-256 | FID: 3.30 -> 2.01 | Improved |
| apply Neon to IMM at 1 step | IMM · ImageNet-256 (T=1) | FID: 7.77 -> 6.67 | Improved |
| jointly tune merge weight w and CFG scale gamma (ablation) | Neon + VAR-d16 · ImageNet-256 | FID: 3.01 -> 2.01 | Improved |
| use synthetic data from other architectures (flow, IMM) for Neon on EDM-VP (ablation) | EDM-VP · CIFAR-10 | FID: 1.97 -> 1.59 (flow); 1.97 -> 1.80 (IMM); 1.97 -> 1.38 (self) | Improved |
| replace synthetic data with CIFAR-10C corrupted real images (ablation) | Neon + EDM-VP · CIFAR-10 | FID | No clear effect |
| use much smaller synthetic dataset (1k samples) for xAR (ablation) | Neon + xAR-L · ImageNet-256 | FID: 1.05 vs 1.02 at 750k | No clear effect |
| use fewer than 90k synthetic samples for VAR-d16 (ablation) | Neon + VAR-d16 · ImageNet-256 | FID | Hurt |
| raise CFG scale used to generate synthetic data to extreme values (>=6) (ablation) | Neon + xAR-B · ImageNet-256 | FID: 1.30-1.31 for gamma in [1,3] then degrades | Hurt |
Closest papers by meaning
- Self-Consuming Generative Models Go MAD (ICLR 2024 · similarity 0.90)
- A Closer Look at Model Collapse: From a Generalization-to-Memorization Perspective (NeurIPS 2025 · similarity 0.86)
- PaGoDA: Progressive Growing of a One-Step Generator from a Low-Resolution Diffusion Teacher (NeurIPS 2024 · similarity 0.85)
- Stretching Each Dollar: Diffusion Training from Scratch on a Micro-Budget (CVPR 2025 · similarity 0.85)
- End-to-End Training for Unified Tokenization and Latent Denoising (NeurIPS 2026 · similarity 0.85)
- Improved Training Technique for Shortcut Models (NeurIPS 2025 · similarity 0.85)
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.
