Research map / Image generation / Autoregressive and latent generation
UltraPixel: Advancing Ultra High-Resolution Image Synthesis to New Peaks
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.
Proposes UltraPixel, a cascade-diffusion T2I framework that uses later-stage low-resolution features as guidance, implicit neural upsampling, and scale-aware normalization to generate 1K-6K images with one set of weights.
Problem
Ultra-high-resolution text-to-image generation suffers from hard semantic planning, detail synthesis difficulties, artifacts from training-free methods, and high training cost; existing HR-specific models are tied to fixed resolutions.
Approach
Build on StableCascade's 42:1 compact latent space; extract later-stage (t=0.05) features from the low-resolution generation as guidance, upsample them continuously with an implicit neural representation, and add scale-aware normalization so one model with ~3% extra parameters handles multiple resolutions.
Key results
- Preferred over PixArt-Sigma in 85.2% (one resolution) and 84.0% of PickScore comparisons, and over DALL-E 3 in 70.0%.
- Generates a 2160x3840 image in 31 seconds versus 111 seconds for PixArt-Sigma, and 4096x4096 in 78 seconds versus 728 seconds for DemoFusion.
- Ablation: t=0.05 LR guidance gives CLIP 33.09 and IS 29.14, outperforming t=0.5 and t=t^H; INR beats bilinear+conv upsampling.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add low-resolution semantic guidance to high-resolution generation (ablation) | UltraPixel · high-resolution image synthesis | visual quality and semantic coherence | Helped |
| extract LR guidance at t=0.05 instead of t=t^H (ablation) | UltraPixel · LR guidance timestep ablation | CLIP: 31.14 -> 33.09 | Helped |
| extract LR guidance at t=0.05 instead of t=0.5 (ablation) | UltraPixel · LR guidance timestep ablation | CLIP: 32.75 -> 33.09 | Helped |
| use t=0.05 LR guidance instead of t=t^H (ablation) | UltraPixel · LR guidance timestep ablation | IS: 25.37 -> 29.14 | Helped |
| use t=0.05 LR guidance instead of t=0.5 (ablation) | UltraPixel · LR guidance timestep ablation | IS: 28.15 -> 29.14 | Helped |
| use INR instead of bilinear interpolation followed by convolutions (ablation) | UltraPixel · 2560² | CLIP: 32.41 -> 32.72 | Helped |
| use INR instead of bilinear interpolation followed by convolutions (ablation) | UltraPixel · 2560² | IS: 26.81 -> 27.62 | Helped |
| add scale-aware normalization to INR (ablation) | UltraPixel · 2560² | IS: 27.62 -> 29.14 | Helped |
| add scale-aware normalization to INR (ablation) | UltraPixel · 4096² | CLIP: 31.93 -> 32.87 | Helped |
| add scale-aware normalization to INR (ablation) | UltraPixel · 4096² | IS: 25.22 -> 27.15 | Helped |
| increase LR feature dimension from 512 to 1024 (ablation) | UltraPixel · trainable parameter ablation | CLIP: 32.78 -> 33.09 | Helped |
| increase LR feature dimension from 512 to 1024 (ablation) | UltraPixel · trainable parameter ablation | IS: 27.43 -> 29.14 | Helped |
Surprising findings
- Guidance extracted at a near-clean timestep (t=0.05) works better than guidance at the synchronized or middle timesteps, since noisy early features are less useful.
Limitations
- Limited quantity and quality of training data constrain realism, especially in complex scenes.
- Training-free baselines and commercial models were evaluated on only 1K images, and FID/IS are computed on downsampled images, making them poorly suited to HR quality.
Experimental setup
Datasets: LAION-Aesthetics SAM self-collected high-quality images
Benchmarks: 1,000 self-collected high-quality images (1024 to 4096)
Metrics: PickScore win rate FID-patch IS-patch FID IS CLIP score latency
Baselines: ScaleCrafter ElasticDiffusion DemoFusion FouriScale PixArt-Sigma DALL-E 3 Midjourney V6 BSRGAN
Builds on: StableCascade LoRA DDIM classifier-free guidance
Closest papers by meaning
- ScaleCrafter: Tuning-free Higher-Resolution Visual Generation with Diffusion Models (ICLR 2024 · similarity 0.90)
- DiffuseHigh: Training-Free Progressive High-Resolution Image Synthesis Through Structure Guidance (AAAI 2025 · similarity 0.89)
- ResMaster: Mastering High-Resolution Image Generation via Structural and Fine-Grained Guidance (AAAI 2025 · similarity 0.89)
- UltraHR-100K: Enhancing UHR Image Synthesis with A Large-Scale High-Quality Dataset (NeurIPS 2025 · similarity 0.89)
- Scalable High-Resolution Pixel-Space Image Synthesis with Hourglass Diffusion Transformers (ICML 2024 · similarity 0.88)
- UltraDiff:Transferring High-Fidelity Priors to Compressed Latent Spaces for High-resolution Image Generation (NeurIPS 2026 · 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.