NeurIPS 2025 best paper: diffusion models start memorizing at ~300 steps per image
Mastbubbles · reddit · 2026-11-14
A Reddit deep-dive explains the NeurIPS 2025 best paper (arXiv 2505.17638), tackling the "diffusion just copies training data" debate.
- Small diffusion models trained on 1,024 face photos: at 100k steps, none of 10,000 samples copied a training photo
- Train the same model to 1.62M steps and 34 of every 100 samples become training photos — memorization is real, but requires heavy overtraining
- Time to generate good images was dataset-size-independent; time until copying began scaled with dataset size at 300 steps per image, so normally-trained models stop far short of memorization
- The author turned it into a free visual explainer using the paper's own faces
Caveats: only 32x32 grayscale faces and small models — nobody has run this on SD-scale models, which have been shown to reproduce training images. The post asks LoRA trainers at how many steps/image their training images start coming back.
More from Multimodal
- Open-source Blender add-on runs NVIDIA Kimodo, ARDY and MotionBricks locally for AI motion generation — bdsqlsz · 2026-11-14
- Sci-fi film KILL/SAVE team releases "fake ads from 2045" made with AI video — killsavecom · 2026-11-14
- Stella & Shango Part 2: Viking-themed AI video short continues — stellaXaiXvip · 2026-11-14
- Fable 5.2 trick: one recolored 3D character populates every spectator NPC — chongdashu · 2026-11-14
- Making Midjourney obey prompts: low --stylize plus --raw, in v8.2 — michaelrabone · 2026-11-14
- User recants: MiniMax H3 may be the best image editing model yet, undeclared — VasaFromParadise · 2026-11-14