Research map / Image generation / Fast diffusion sampling
NitroFusion: High-Fidelity Single-Step Diffusion through Dynamic Adversarial Training
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.
Distills multi-step SDXL-family teachers (DMD2, Hyper-SDXL) into a one-step student using an adversarial loss with a large pool of specialized discriminator heads on a frozen teacher UNet encoder, refreshed periodically and trained with multi-scale, dual-objective GAN losses.
Problem
One-step diffusion generation typically loses quality (blur, missing detail, artifacts) relative to multi-step models, and adversarial distillation suffers from discriminator overconfidence, instability and diversity collapse.
Approach
A dynamic adversarial framework: a large pool of lightweight discriminator heads attached to a frozen teacher UNet encoder, specialized per noise level and per global/local/conditional/unconditional task, with random sampling per iteration, ~1% head re-initialization per iteration, and bottom-up training so the same model can refine across 1-4 steps.
Key results
- 1-step NitroSD-Realism reports Aesthetic Score 5.56 and ImageReward 0.856 vs DMD2 1-step 5.47 and 0.825 on COCO-5K.
- 1-step NitroSD-Vibrant achieves FID 38.49 and ImageReward 0.991, with FID worse than DMD2 (23.91) and Hyper-SDXL (36.77) reported in the same table.
- In a two-choice user study, NitroSD-Vibrant 1-step was preferred over all compared models including SDXL 25-step (63.3% vs SDXL 50% reference line in Figure 5).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add multi-scale dual-objective GAN training (ablation) | NitroFusion · Image generation | fine-grained detail and image coherence | Helped |
| Add discriminator pool refresh (ablation) | NitroFusion · Image generation | image quality and sharpness | Helped |
| Add a dynamic discriminator pool (ablation) | NitroFusion · Image generation | image sharpness | Helped |
| Use NitroSD-Realism instead of DMD2 | COCO-5K validation, 1 step · COCO-5K validation | Aesthetic Score: 5.47 -> 5.56 | Helped |
| Use NitroSD-Realism instead of DMD2 | COCO-5K validation, 1 step · COCO-5K validation | FID: 23.91 -> 25.61 | Hurt |
| Use NitroSD-Realism instead of DMD2 | COCO-5K validation, 4 steps · COCO-5K validation | Image Reward: 0.880 -> 0.945 | Helped |
| Use NitroSD-Vibrant instead of Hyper-SDXL | COCO-5K validation, 1 step · COCO-5K validation | Aesthetic Score: 6.00 -> 5.92 | Hurt |
| Use NitroSD-Vibrant instead of Hyper-SDXL | COCO-5K validation, 1 step · COCO-5K validation | Image Reward: 1.169 -> 0.991 | Hurt |
Surprising findings
- Quantitatively, FID is not uniformly improved: NitroSD-Vibrant's 1-step FID (38.49) is worse than DMD2's (23.91), and NitroSD-Realism's 1-step CLIP score is not best among compared models.
Limitations
- Quantitative FID results are mixed; the one-step Vibrant variant has notably worse FID than several baselines.
- Evaluation relies on automatic metrics plus a two-choice user study, with no reported variance or significance testing.
Experimental setup
Datasets: Pick-a-Pic LAION
Benchmarks: COCO-5K validation
Metrics: CLIP score FID Aesthetic Score ImageReward
Baselines: SDXL-Turbo SDXL-Lightning Hyper-SDXL DMD2 SDXL-Base
Builds on: Adversarial Diffusion Distillation SDXL-Lightning DMD2 StyleGAN-T Hyper-SDXL
Closest papers by meaning
- Sample-specific Noise Injection for Diffusion-based Adversarial Purification (ICML 2025 · similarity 0.87)
- ACT-Diffusion: Efficient Adversarial Consistency Training for One-step Diffusion Models (CVPR 2024 · similarity 0.87)
- Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (ICML 2025 · similarity 0.87)
- You Only Sample Once: Taming One-Step Text-to-Image Synthesis by Self-Cooperative Diffusion GANs (ICLR 2025 · similarity 0.87)
- Divide and Conquer: Heterogeneous Noise Integration for Diffusion-based Adversarial Purification (CVPR 2025 · similarity 0.87)
- OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.87)
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.