Research map / Image generation / Fast diffusion sampling
Universal Inverse Distillation for Matching Models with Real-Data Supervision (No GANs)
Nikita Maksimovich Kornilov, David Li, Tikhon Mavrin, Aleksei Leonov, Nikita Gushchin, Evgeny Burnaev, Iaroslav Sergeevich Koshelev, Alexander Korotin
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 |
|---|---|---|---|
| incorporate real data into distillation via RealUID coefficients with beta/alpha near 0.98 or 1.02 (ablation) | UID (FGM) flow distillation · CIFAR-10 unconditional | FID: 2.58 -> 2.23 | Improved |
| incorporate real data via RealUID with beta/alpha near 0.98 (ablation) | UID (FGM) flow distillation · CIFAR-10 conditional | FID: 2.21 -> 2.02 | Improved |
| use real data with beta/alpha = 1 or values outside [0.98, 1.02] (ablation) | RealUID · CIFAR-10 | FID: unconditional baseline 2.58; beta/alpha=0.96 gives 2.66-2.97 | Hurt |
| add GAN adversarial loss with discriminator for real data (ablation) | UID (FGM) · CIFAR-10 unconditional | FID: 2.58 -> 2.29 | Improved |
| increase GAN loss weights to 5/15 (ablation) | UID + GAN · CIFAR-10 unconditional | FID: 2.29 -> 2.54 | Hurt |
| use RealUID instead of GAN for real data (ablation) | UID (FGM) · CIFAR-10 conditional | FID: 2.12 (GAN) -> 2.02 (RealUID) | Improved |
| add real-data coefficients to speed convergence | RealUID · CIFAR-10 | iterations to saturation: ~300k -> ~100k iterations | Improved |
| add fine-tuning stage with new alpha_FT, beta_FT from best checkpoint | RealUID · CIFAR-10 unconditional | FID: 2.23 -> 1.98 | Improved |
| add fine-tuning stage | RealUID · CIFAR-10 conditional | FID: 2.02 -> 1.87 | Improved |
| use RealUID with fine-tuning instead of GAN with fine-tuning (ablation) | UID · CIFAR-10 unconditional | FID: 2.10 -> 1.98 | Improved |
| use RealUID with fine-tuning instead of GAN with fine-tuning (ablation) | UID · CIFAR-10 conditional | FID: 1.88 -> 1.87 | No clear effect |
| add real data, student vs teacher | RealUID · CIFAR-10 | FID: Teacher 3.57 -> 1.98 (uncond); 5.56 -> 1.87 (cond) | Improved |
Closest papers by meaning
- Path-Guided Flow Matching for Dataset Distillation (NeurIPS 2026 · similarity 0.88)
- Self-Corrected Flow Distillation for Consistent One-Step and Few-Step Image Generation (AAAI 2025 · similarity 0.88)
- Inductive Moment Matching (ICML 2025 · similarity 0.87)
- Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (ICML 2025 · similarity 0.87)
- Efficient Dataset Distillation for Pre-Trained Self-Supervised Models via Statistical Flow Matching (NeurIPS 2026 · similarity 0.87)
- Indirect Gradient Matching for Adversarial Robust Distillation (ICLR 2025 · similarity 0.86)
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.
