Research map / Image generation / Text-to-image diffusion
SwiftBrush v2: Make Your One-step Diffusion Model Better Than Its Teacher
Trung Tuan Dao, Thuan Hoang Nguyen, Thanh Van Le, Duc H Vu, Khoi Nguyen, Cuong Pham, Anh T Tran
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 |
|---|---|---|---|
| initialize SwiftBrush student with SD Turbo pretrained weights (ablation) | SwiftBrush · MS COCO-2014 30K | FID: 15.46 -> 14.59 | Improved |
| add 2M LAION prompts to the training set (ablation) | SwiftBrush (Turbo init.) · MS COCO-2014 30K | FID: 14.59 -> 11.27 | Improved |
| add clamped CLIP loss in LoRA efficient training (ablation) | SwiftBrush v2 · MS COCO-2014 30K | FID: 13.21 -> 11.70 | Improved |
| add unclamped (naive) CLIP loss (ablation) | SwiftBrush v2 · MS COCO-2014 30K | FID: 14.59 -> 21.32 | Hurt |
| clamp CLIP loss with ReLU margin instead of naive CLIP (ablation) | SwiftBrush v2 · MS COCO-2014 30K | FID: 21.32 -> 13.19 | Improved |
| add weight scheduler for clamped CLIP loss (ablation) | SwiftBrush v2 · MS COCO-2014 30K | FID: 13.19 -> 11.70 | Improved |
| merge fully trained Model A and LoRA Model B by weight interpolation (ablation) | SwiftBrush v2 · MS COCO-2014 30K | FID: 11.27 / 11.02 -> 8.77 | Improved |
| add extra image regularization on minimal real data (ablation) | SwiftBrush v2 merged model · MS COCO-2014 30K | FID: 8.77 -> 8.14 | Improved |
| add LoRA and TinyVAE efficient training scheme (ablation) | SwiftBrush v2 with CLIP loss · training cost | GPU days: 7.8 -> 4.3 | Improved |
| scale training prompts from 1.3M to 3.3M (ablation) | SwiftBrush (Turbo init.) · MS COCO-2014 30K | Recall: 0.55 -> 0.54 | Mixed |
| initialize from SwiftBrush instead of SD Turbo (ablation) | SwiftBrush v2 · MS COCO-2014 30K | FID: 8.77 -> 11.69 | Hurt |
| add HPSv2 loss to efficient training (ablation) | Model B · MS COCO-2014 30K and HPSv2 | FID / HPSv2 Photo: FID 11.02 -> 11.15; HPSv2 Photo 26.71 -> 28.00 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- SwiftBrush — SwiftBrush: One-Step Text-to-Image Diffusion Model with Variational Score Distillation (CVPR 2024) · better on 1 of 1 shared result
- InstaFlow-0.9B — InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation (ICLR 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SwiftBrush: One-Step Text-to-Image Diffusion Model with Variational Score Distillation (CVPR 2024 · similarity 0.93)
- Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (ICML 2025 · similarity 0.89)
- On the Scalability of Diffusion-based Text-to-Image Generation (CVPR 2024 · similarity 0.88)
- You Only Sample Once: Taming One-Step Text-to-Image Synthesis by Self-Cooperative Diffusion GANs (ICLR 2025 · similarity 0.88)
- InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation (ICLR 2024 · similarity 0.88)
- Multistep Distillation of Diffusion Models via Moment Matching (NeurIPS 2024 · similarity 0.88)
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.
