Research map / Image generation / Fast diffusion sampling
Uni-Instruct: One-step Diffusion Model through Unified Diffusion Divergence Instruction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Uni-Instruct, a theory-driven framework that unifies more than 10 one-step diffusion distillation methods via a diffusion expansion of f-divergences and a tractable gradient-equivalent loss, achieving state-of-the-art one-step FIDs on CIFAR10 and ImageNet 64x64.
Problem
One-step diffusion distillation methods (KL-based like Diff-Instruct/DMD and score-based like SIM/SiD) appear separate in theory, and existing approaches have mode-collapse or slow-convergence issues; the paper asks whether they can be unified and whether this improves one-step generators.
Approach
Expands the f-divergence between the one-step model and teacher along a diffusion process into an integral of score-difference terms weighted by curvature coefficients, then derives a tractable gradient-equivalent loss (combining Diff-Instruct and SIM gradients) with a GAN-based density-ratio estimate; special cases recover DI, SIM, SiD, DMD, f-distill, SDS and VSD.
Key results
- On CIFAR-10 conditional generation, Uni-Instruct (FKL, longer training) reaches one-step FID 1.38, versus 1.40 for Si^2D A and 1.44 for SiDA.
- On ImageNet 64x64 conditional generation, the paper reports a one-step FID of 1.02, versus 2.35 for its 79-step teacher.
- On CIFAR-10 unconditional generation, the JKL variant reaches FID 1.46 in the ablation table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use JKL divergence instead of χ², RKL, or FKL (ablation) | Uni-Instruct trained from scratch · CIFAR10 unconditional | FID: 1.52 -> 1.46 | Helped |
| Add GAN loss to Uni-Instruct RKL (ablation) | Uni-Instruct RKL · CIFAR10 unconditional | FID: 1.92 -> 1.52 | Helped |
| Add GAN loss to Uni-Instruct FKL (ablation) | Uni-Instruct FKL · CIFAR10 unconditional | FID: 1.88 -> 1.52 | Helped |
| Add a SiD model as initialization (ablation) | Uni-Instruct FKL · CIFAR10 unconditional | FID: 1.52 -> 1.48 | Helped |
| Add a SiD model as initialization (ablation) | Uni-Instruct RKL · CIFAR10 unconditional | FID: 1.52 -> 1.50 | Helped |
| Add a SiD model as initialization (ablation) | Uni-Instruct JKL · CIFAR10 unconditional | FID: 1.46 -> 1.50 | No clear effect |
| Use JS divergence (ablation) | Uni-Instruct · CIFAR10 unconditional | FID: 1.46 -> 5.23 | Hurt |
| Remove the divergence loss and retain GAN loss (ablation) | Uni-Instruct · CIFAR10 unconditional | FID: 1.46 -> 8.21 | Hurt |
| Replace JKL with χ² divergence (ablation) | Uni-Instruct · CIFAR10 unconditional | FID: 1.46 -> 4.37 | Hurt |
| Resume RKL training with FKL for longer training | Uni-Instruct · CIFAR10 conditional | FID: 1.43 -> 1.38 | Helped |
| Replace SiDA with Uni-Instruct FKL | One-step generator · CIFAR10 unconditional | FD_dino: 132.72 -> 129.30 | Helped |
| Replace SiDA with Uni-Instruct FKL | One-step generator · ImageNet 64×64 | IS: 59.28 -> 64.11 | Helped |
Surprising findings
- Uni-Instruct one-step generator (1.02 FID) outperforms its own 79-step teacher on ImageNet 64x64.
- The JS divergence variant yields unsuccessful distillation due to numerical instability of its weightings.
Limitations
- Gradient derivation ignores the dependence of x on theta in the score-difference term, following SIM for simplicity.
- No error bars reported due to computational overhead.
- Text-to-3D results are only qualitative and described as slightly better than ProlificDreamer.
Experimental setup
Datasets: CIFAR-10 ImageNet 64x64
Benchmarks: CIFAR-10 ImageNet 64x64
Metrics: FID sFID FD_DINO Inception Score Precision Recall
Baselines: Diff-Instruct SIM CTM SiD SiDA Si^2D A f-distill DMD DDPM iDDPM SDS VSD
Builds on: Diff-Instruct Score Implicit Matching SiD f-distill SiDA
Closest papers by meaning
- Diffusion Model is Effectively Its Own Teacher (CVPR 2025 · similarity 0.91)
- One-Step Diffusion Distillation through Score Implicit Matching (NeurIPS 2024 · similarity 0.91)
- One-step Diffusion with Distribution Matching Distillation (CVPR 2024 · similarity 0.90)
- Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation (ICML 2024 · similarity 0.90)
- Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (ICML 2025 · similarity 0.90)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.89)
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.