Research map / Image generation / Super-resolution
Generative Powers of Ten
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.
Generative Powers of Ten uses a pretrained text-to-image diffusion model (Imagen) with joint multi-scale diffusion sampling and Laplacian multi-resolution blending to generate zoom sequences that stay consistent across scales.
Problem
Generating consistent content across extreme semantic zoom levels of a scene, where finer zooms reveal new structures that super-resolution and outpainting cannot reliably synthesize.
Approach
Run parallel diffusion sampling at all zoom levels, each guided by its own text prompt, and fuse their clean-image estimates into a shared zoom stack via frequency-band (Laplacian pyramid) blending, with shared noise across scales, so that all levels are jointly consistent during sampling.
Key results
- Ten example zoom sequences are generated, with prompt sequences of 6 to 16 prompts, shown qualitatively.
- Qualitatively, the method produces more consistent zoom sequences than progressive outpainting and progressive super-resolution from Stable Diffusion, which accumulate errors or fail to synthesize new zoomed-in objects.
- Ablations show independent sampling, iterative update, shared-noise removal, and naive blending all produce less consistent or blurrier results.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use multi-scale joint sampling instead of independent sampling at each scale | Generative Powers of Ten · Text-conditioned zoom sequences | cross-scale consistency | Helped |
| Use multi-scale joint sampling instead of progressive outpainting | Generative Powers of Ten · Text-conditioned zoom sequences | cross-scale consistency | Helped |
| Use multi-scale joint sampling instead of progressive super-resolution | Generative Powers of Ten · Text-conditioned zoom sequences | fine-scale content generation | Helped |
| Update all scales jointly rather than iteratively updating each level separately (ablation) | Generative Powers of Ten · Text-conditioned zoom sequences | cross-scale consistency | Helped |
| Use shared noise across zoom levels rather than independently sampled noise (ablation) | Generative Powers of Ten · Text-conditioned zoom sequences | image sharpness | Helped |
| Use multi-resolution blending rather than naïve blending (ablation) | Generative Powers of Ten · Text-conditioned zoom sequences | image sharpness | Helped |
Limitations
- Finding prompt sets that agree across fixed scales and are generable by the model is a significant challenge and is manual.
- Evaluation is qualitative only, with no quantitative consistency metrics.
- Relies on an internal Imagen model that is not publicly available.
Experimental setup
Baselines: Stable Diffusion outpainting Stable Diffusion upscale super-resolution independent sampling naive blending (MultiDiffusion-style) iterative update
Builds on: Imagen MultiDiffusion DDPM classifier-free guidance
Closest papers by meaning
- ZoomLDM: Latent Diffusion Model for Multi-scale Image Generation (CVPR 2025 · similarity 0.87)
- Generative Photography: Scene-Consistent Camera Control for Realistic Text-to-Image Synthesis (CVPR 2025 · similarity 0.87)
- Taming Stable Diffusion for Text to 360 Panorama Image Generation (CVPR 2024 · similarity 0.87)
- The Power of Context: How Multimodality Improves Image Super-Resolution (CVPR 2025 · similarity 0.86)
- ScaleCrafter: Tuning-free Higher-Resolution Visual Generation with Diffusion Models (ICLR 2024 · similarity 0.86)
- Image Regeneration: Evaluating Text-to-Image Model via Generating Identical Image with Multimodal Large Language Models (AAAI 2025 · similarity 0.86)
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.