Research map / Image generation / Text-to-image diffusion
Guided Score identity Distillation for Data-Free One-Step 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.
Extends Score identity Distillation (SiD) with long-short classifier-free guidance (LSG) to distill Stable Diffusion 1.5 and 2.1-base into data-free one-step text-to-image generators, reaching a record data-free COCO-2014 zero-shot FID of 8.15.
Problem
Slow iterative sampling of text-to-image diffusion models; existing one-step distillation methods usually need real images or extra regression/adversarial losses, and SiD had not been adapted to classifier-free-guided latent T2I diffusion models.
Approach
Applies classifier-free guidance not only to the pretrained score network but also to the training and evaluation of the fake score network within SiD's model-based explicit score matching loss (long-short guidance), with the generator trained only on its own synthetic images.
Key results
- SiD-LSG distilled from SD1.5 with LSG scale 1.5 and doubled training reaches a zero-shot COCO-2014 FID of 8.15 with CLIP 0.304, a record among data-free one-step diffusion distillation methods.
- With guidance scale 2, SiD-LSG attains FID 9.56 (SD1.5, CLIP 0.313) and 10.97 (SD2.1-base, CLIP ~0.318).
- On HPSv2, SiD-LSG outperforms SwiftBrush and other baselines in most categories, except ADD which uses real data and adversarial loss.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add classifier-free guidance rather than using no CFG (ablation) | SiD distillation of Stable Diffusion 1.5 · COCO-2014 validation | FID: 15.49 -> 10.01 | Helped |
| add classifier-free guidance rather than using no CFG (ablation) | SiD distillation of Stable Diffusion 1.5 · COCO-2014 validation | CLIP: 0.269 -> 0.297 | Helped |
| increase the common LSG scale from 1.5 to 4.5 (ablation) | SiD-LSG with Stable Diffusion 1.5 · COCO-2014 validation | FID: 8.71 -> 16.59 | Hurt |
| increase the common LSG scale from 1.5 to 4.5 (ablation) | SiD-LSG with Stable Diffusion 1.5 · COCO-2014 validation | CLIP: 0.302 -> 0.317 | Helped |
| double training duration from 10M to 20M fake images (ablation) | SiD-LSG with κ=1.5 and Stable Diffusion 1.5 · COCO-2014 validation | FID: 8.71 -> 8.15 | Helped |
| switch prompts from Aesthetics6+ to Aesthetics6.25+ (ablation) | SiD-LSG with κ=2 and Stable Diffusion 1.5 · COCO-2014 validation | FID: 9.56 -> 9.21 | Helped |
| switch prompts from Aesthetics6+ to Aesthetics6.25+ (ablation) | SiD-LSG with κ=2 and Stable Diffusion 1.5 · COCO-2014 validation | CLIP: 0.313 -> 0.311 | Hurt |
| vary batch size and learning rate (batch size 8192, learning rate 4e-6; or batch size 64, learning rate 1e-6) (ablation) | SiD-LSG with κ=1.5 and Stable Diffusion 1.5 · COCO-2014 validation | FID | No clear effect |
| increase fake-score training guidance alone to κ1=2 (ablation) | SiD distillation of Stable Diffusion 1.5 · COCO-2014 validation | CLIP: 0.269 -> 0.306 | Helped |
| use the recommended equal-scale LSG rather than long CFG (κ4=2) (ablation) | SiD distillation of Stable Diffusion 1.5 · COCO-2014 validation | FID: 10.01 -> 10.44 | Mixed |
| use the recommended equal-scale LSG rather than long CFG (κ4=2) (ablation) | SiD distillation of Stable Diffusion 1.5 · COCO-2014 validation | CLIP: 0.297 -> 0.304 | Helped |
| increase batch size or change training prompts (ablation) | SiD-LSG with Stable Diffusion 1.5 · COCO-2014 validation | FID | No clear effect |
Surprising findings
- Applying CFG to the fake score network during training (not only the teacher) yields the best FID-CLIP trade-off, with the simplest variant (only kappa_1>1) already matching the best long or short strategy alone.
- Lower LSG scales give better FID but worse CLIP, and the one-step student on SD1.5 can surpass the 50-step teacher's FID of 8.78 while still lagging its CLIP score.
Limitations
- SiD-LSG reaches a performance plateau at larger LSG scales and still lags the teacher's CLIP score at the best-FID setting.
- FP16 training is less stable and gives worse best FID/CLIP than FP32; requires gradient clipping.
- Compute cost: roughly 20% more time per iteration and about 10% more peak memory than long or short strategies; requires gradient backpropagation through score networks.
Experimental setup
Datasets: Aesthetics6+ prompts COCO-2014 validation prompts
Benchmarks: COCO-2014 validation set (zero-shot FID-30K) COCO-2014 CLIP score HPSv2 COCO-2014 Precision and Recall
Metrics: FID CLIP score HPSv2 precision recall inference time
Baselines: ADD (SD-Turbo) SDXL-Turbo ADD DMD UFOGen InstaFlow SwiftBrush BOOT LCM-LoRA UniPC DPM++ Guided Distillation Muse
Builds on: Score identity Distillation (SiD) Distribution Matching Distillation (DMD) classifier-free guidance EDM
Closest papers by meaning
- Noise-free Score Distillation (ICLR 2024 · similarity 0.92)
- Diffusion Self-Distillation for Zero-Shot Customized Image Generation (CVPR 2025 · similarity 0.91)
- SwiftBrush: One-Step Text-to-Image Diffusion Model with Variational Score Distillation (CVPR 2024 · similarity 0.90)
- Plug-and-Play Diffusion Distillation (CVPR 2024 · similarity 0.90)
- Rethinking Score Distillation as a Bridge Between Image Distributions (NeurIPS 2024 · similarity 0.90)
- Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation (ICML 2024 · similarity 0.90)
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.