Research map / Image generation / Fast diffusion sampling
InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation
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.
Applies text-conditioned Rectified Flow reflow to Stable Diffusion, then distills into a one-step text-to-image model (InstaFlow) trained with supervised learning in 199 A100 GPU days.
Problem
Multi-step diffusion text-to-image sampling is slow; direct distillation of Stable Diffusion into a one-step model fails due to curved probability-flow trajectories and poor noise-image coupling.
Approach
Introduce a text-conditioned reflow procedure from Rectified Flow to straighten SD's probability flow ODE and refine noise-image couplings, then distill the straightened 2-Rectified Flow into a single Euler step with LPIPS/L2 losses, optionally using a stacked two-U-Net student at 1.7B parameters.
Key results
- InstaFlow-0.9B achieves FID-5k 23.4 in 0.09s on MS COCO 2017, versus 37.2 for one-step Progressive Distillation-SD.
- InstaFlow-1.7B achieves FID-5k 22.4 in 0.12s; InstaFlow-0.9B achieves FID-30k 13.10 on MS COCO 2014 in 0.09s, versus StyleGAN-T at 13.90 in 0.1s.
- Direct distillation of SD 1.4 gives FID-5k 40.9 in one step, while distillation after 2-Rectified Flow reflow gives 31.0 under the same preliminary setup.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add text-conditioned reflow before one-step distillation | Stable Diffusion 1.4 one-step distillation · MS COCO 2017-5k | FID: 40.9 -> 31.0 | Helped |
| Add text-conditioned reflow before one-step distillation | Stable Diffusion 1.4 one-step distillation · MS COCO 2014-30k | FID: 34.6 -> 20.0 | Helped |
| Increase the model size from 0.9B to 1.7B parameters | InstaFlow · MS COCO 2017-5k | FID: 23.4 -> 22.4 | Helped |
| Increase the model size from 0.9B to 1.7B parameters | InstaFlow · MS COCO 2014-30k | FID: 13.10 -> 11.83 | Helped |
| Distill from 3-Rectified Flow rather than 2-Rectified Flow | (Pre) Rectified Flow+Distill (U-Net) · MS COCO 2017-5k | FID: 31.0 -> 29.3 | Mixed |
| Distill from 3-Rectified Flow rather than 2-Rectified Flow | (Pre) Rectified Flow+Distill (Stacked U-Net) · MS COCO 2017-5k | FID: 24.6 -> 26.3 | Hurt |
| Raise guidance scale α | 2-Rectified Flow · MS COCO 2017 | FID-5k | Hurt |
| Raise guidance scale α | 2-Rectified Flow · MS COCO 2017 | CLIP | Helped |
| Remove Downsample Blocks 1 (ablation) | Concatenated two-U-Net distillation network · Text-to-image generation | generation quality | Hurt |
| Remove Upsample Blocks 1 (ablation) | Concatenated two-U-Net distillation network · Text-to-image generation | generation quality | Hurt |
| Remove In+Out Block (ablation) | Concatenated two-U-Net distillation network · Text-to-image generation | generation quality | No clear effect |
| Remove Downsample Blocks 2 (ablation) | Concatenated two-U-Net distillation network · Text-to-image generation | generation quality | Hurt |
Surprising findings
- Direct one-step distillation from Stable Diffusion fails (FID 40.9 or worse across a grid of learning rates and weight decays), while reflow before distillation makes it work.
- Guidance scale increases for 2-Rectified Flow raise FID while improving CLIP score, showing a quality-alignment trade-off.
Limitations
- Struggles with complex compositions in text prompts, per the failure case discussion.
- Reflow and distillation training was limited in duration; 2-Rectified Flow did not fully converge according to the authors.
- Training cost (199 A100 GPU days) remains substantial despite being far below GAN-based large-scale models.
Experimental setup
Datasets: LAION-2B-en MS COCO 2017 MS COCO 2014
Benchmarks: MS COCO 2017 5k MS COCO 2014 30k
Metrics: FID CLIP score inference time
Baselines: Progressive Distillation-SD StyleGAN-T GigaGAN SD 1.5 (25-step DPM-Solver) SD 1.4+Distill Parti Imagen GLIDE LDM
Builds on: Rectified Flow Stable Diffusion Progressive Distillation DPM-Solver LPIPS
Closest papers by meaning
- Scaling Rectified Flow Transformers for High-Resolution Image Synthesis (ICML 2024 · similarity 0.90)
- InitNO: Boosting Text-to-Image Diffusion Models via Initial Noise Optimization (CVPR 2024 · similarity 0.90)
- SwiftBrush: One-Step Text-to-Image Diffusion Model with Variational Score Distillation (CVPR 2024 · similarity 0.90)
- Text-to-Image Rectified Flow as Plug-and-Play Priors (ICLR 2025 · similarity 0.89)
- Rectified Diffusion: Straightness Is Not Your Need in Rectified Flow (ICLR 2025 · similarity 0.89)
- Fast constrained sampling in pre-trained diffusion models (NeurIPS 2025 · 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.