IDU: Unified Unlearning and One-Step Distillation for Flow and Diffusion Models
Aleksei Leonov · hf · 2026-10-06
A Hugging Face paper introduces Inverse Distillation Unlearning (IDU), reportedly the first unified framework combining unlearning and distillation for unconditional flow-matching and score-based models: distill a multi-step teacher into an efficient one-step student while suppressing outputs tied to a forget set.
Approach: formulate distillation as a min-max objective over a data distribution, represent it as a mixture of the forget-set and generated distributions, and compare against the teacher's training distribution so only retained data is recovered at optimum. Requires only a pretrained full-data teacher and forget-set data—no access to retained samples, extra feature extractors, or classifiers.
Results: on MNIST and CIFAR-10 under both settings, IDU substantially cuts generation of forgotten classes while preserving quality on retained classes.
More from Multimodal
- Prompt Share: Holographic Data Projection for Sci-Fi AI Images — azed_ai · 2026-10-06
- Open-source huashu-art-motion skill animates 35 art styles via coding agents — AlchainHust · 2026-10-06
- One Prompt Runs Claude Opus and Seedance Together for Video — Itchy-Car-904 · 2026-10-06
- MiniMax H3 Local Benchmark: How Long for 17s 720p Video on RTX 5080? — off_rs · 2026-10-06
- A Krea 2 realism workflow: Chroma for composition, LoRA for character consistency — Wide_Director_8897 · 2026-10-06
- AI-generated anime stage clip wows Reddit with smoke and lightning vibes — Amazing_Skill_6080 · 2026-10-06