Research map / Image generation / Text-to-image diffusion
ElasticDiffusion: Training-free Arbitrary Size Image Generation through Global-Local Content Separation
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.
Training-free decoding method (ElasticDiffusion) that lets a pretrained SD1.4 generate arbitrary-size images by separating local (patch-wise unconditional score) and global (downsampled class-direction score) signals.
Problem
Pretrained text-to-image diffusion models are confined to their training resolutions and aspect ratios; naive resolution changes cause repetition and structural incoherence, while fine-tuning or patch-averaging methods are costly.
Approach
Decouple the diffusion trajectory: estimate the unconditional score on local patches with implicit context overlap (no averaging), and estimate the class-direction score from a reference latent at the training resolution, upsampled to the target size, with iterative resampling and reduced-resolution guidance to suppress artifacts.
Key results
- At 1024x1024, ElasticDiffusion (SD1.4) reaches FID 228.87 and CLIP 23.74 on CelebA-HQ versus SDXL's 230.21 and 24.62, using about 8.6GB versus 27.5GB memory.
- At 256x256 on CelebA-HQ, ElasticDiffusion improves FID to 235.23 from SD1.4's 258.43 with CLIP 23.88 vs 20.14.
- Ablation on 500 images: removing resampling raises FID from 133.67 to 150.01; removing RRG raises it to 150.64.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add iterative resampling (ablation) | ElasticDiffusion · Ablation, 500 images | FID: 150.01 -> 133.67 | Helped |
| add iterative resampling (ablation) | ElasticDiffusion · Ablation, 500 images | CLIP: 23.82 -> 25.82 | Helped |
| add reduced-resolution guidance (ablation) | ElasticDiffusion · Ablation, 500 images | FID: 150.64 -> 133.67 | Helped |
| add reduced-resolution guidance (ablation) | ElasticDiffusion · Ablation, 500 images | CLIP: 24.34 -> 25.82 | Helped |
| use implicit overlap instead of explicit overlap (ablation) | ElasticDiffusion · Ablation, 500 images | FID: 141.42 -> 133.67 | Helped |
| use implicit overlap instead of explicit overlap (ablation) | ElasticDiffusion · Ablation, 500 images | CLIP: 25.82 -> 25.82 | No clear effect |
Surprising findings
- Stable Diffusion's CLIP score increases with resolution even as FID worsens, which the authors attribute to repetitive textures aligning with prompts.
- SDXL produces noise-like outputs at 256x256 despite strong results at its native resolution.
Limitations
- Produces blurrier outputs due to global-local guidance and resampling.
- Less effective at extended sizes beyond the training resolution; artifacts beyond about 4x the training size.
- Global content is estimated at the training resolution, limiting extrapolation.
Experimental setup
Datasets: CelebA-HQ (Multi-Modal) LAION-COCO
Benchmarks: CelebA-HQ LAION-COCO
Metrics: FID CLIPscore
Baselines: Stable Diffusion 1.4 MultiDiffusion SDXL Any-Size-Diffusion
Builds on: Stable Diffusion MultiDiffusion classifier-free guidance DDIM
Closest papers by meaning
- Any-Size-Diffusion: Toward Efficient Text-Driven Synthesis for Any-Size HD Images (AAAI 2024 · similarity 0.87)
- Move Anything with Layered Scene Diffusion (CVPR 2024 · similarity 0.87)
- Ambient Diffusion Omni: Training Good Models with Bad Data (NeurIPS 2025 · similarity 0.87)
- FAM Diffusion: Frequency and Attention Modulation for High-Resolution Image Generation with Stable Diffusion (CVPR 2025 · similarity 0.87)
- DCTdiff: Intriguing Properties of Image Generative Modeling in the DCT Space (ICML 2025 · similarity 0.87)
- RealCompo: Balancing Realism and Compositionality Improves Text-to-Image Diffusion Models (NeurIPS 2024 · 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.