Research map / Image generation / Text-to-image diffusion
Ultra-Resolution Adaptation with Ease
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 URAE, a set of guidelines for adapting FLUX.1-dev to 2K/4K generation with few samples, using synthetic teacher data, minor-singular-component tuning when synthetic data are absent, and disabling CFG during adaptation.
Problem
Ultra-resolution (2K/4K) text-to-image adaptation of diffusion models normally requires massive high-quality datasets and large compute, and full fine-tuning of large DiT backbones is memory-intensive.
Approach
Combine three guidelines: train on synthetic data from a strong teacher (FLUX1.1 [Pro] Ultra) when available, otherwise tune the minor (smallest-singular-value) components of weight matrices instead of LoRA's major components, and set the guidance scale to 1 during adaptation for guidance-distilled models like FLUX.
Key results
- URAE on FLUX-1.dev* at 2K achieves FID 29.44 and PickScore 23.15 on HPD prompts, versus 34.86 and 22.68 for the FLUX-1.dev* baseline and 43.78 for FLUX-1.dev.
- With 3K synthetic samples and 2K iterations, URAE improves DPG Bench from 80.15 (FLUX-1.dev*) to 83.83, versus 84.76 for FLUX1.1 [Pro] Ultra.
- At 4K, the minor-component variant (Minor-4K) achieves the top user-study preference (about 53% overall quality) and MAN-IQA of 0.3999 versus 0.2935 for PixArt-Sigma-XL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply URAE to FLUX-1.dev* | FLUX-1.dev* · HPD Prompt, 2048 × 2048 | FID: 34.86 -> 29.44 | Helped |
| Apply URAE to FLUX-1.dev* | FLUX-1.dev* · DPG Prompt, 2048 × 2048 | DPG Bench: 80.15 -> 83.83 | Helped |
| Apply URAE to I-Max | I-Max · DPG Prompt, 2048 × 2048 | DPG Bench: 83.92 -> 87.88 | Helped |
| Apply URAE to SDEdit | SDEdit · HPD Prompt, 2048 × 2048 | FID: 35.59 -> 34.07 | Helped |
| Apply URAE to I-Max | I-Max · HPD Prompt, 2048 × 2048 | FID: 33.66 -> 32.24 | Helped |
| Enable CFG during training (ablation) | Major-component tuning with synthetic data · HPD prompts, 2048 × 2048 | FID: 29.44 -> 76.07 | Hurt |
| Enable CFG during training (ablation) | Minor-component tuning with synthetic data · HPD prompts, 2048 × 2048 | FID: 27.90 -> 65.34 | Hurt |
| Enable CFG during training (ablation) | Major-component tuning with real data · HPD prompts, 2048 × 2048 | FID: 31.39 -> 133.68 | Hurt |
| Enable CFG during training (ablation) | Minor-component tuning with real data · HPD prompts, 2048 × 2048 | FID: 32.09 -> 133.32 | Hurt |
| Use synthetic rather than real training data (ablation) | Major-component tuning without CFG · HPD prompts, 2048 × 2048 | FID: 31.39 -> 29.44 | Helped |
| Use synthetic rather than real training data (ablation) | Minor-component tuning without CFG · HPD prompts, 2048 × 2048 | FID: 32.09 -> 27.90 | Helped |
| Tune minor rather than major components (ablation) | Synthetic-data training without CFG · HPD prompts, 2048 × 2048 | FID: 29.44 -> 27.90 | Helped |
Surprising findings
- Training with CFG enabled (guidance scale > 1) causes severe degradation, e.g. FID rising from 29.44 to 76.07 with synthetic data.
- Tuning minor singular components outperforms LoRA when no synthetic data are available, while LoRA (major components) is preferred with synthetic data at 2K.
Limitations
- Inference-time efficiency is not improved; the approach does not address the slow high-resolution generation.
- 2K results depend on access to a closed-weight teacher (FLUX1.1 [Pro] Ultra) for synthetic data; theory relies on infinite-width linearization assumptions.
- No well-defined 4K benchmark; 4K evaluation relies on a user study and GPT-4o judgments.
Experimental setup
Datasets: FLUX1.1 [Pro] Ultra synthetic images (3K) LAION-5B subset (30K images >=4K resolution) HPD DPG COCO30K
Benchmarks: HPD prompts (2048x2048) DPG Bench COCO30K prompts (AI preference) 4K user study
Metrics: FID LPIPS MAN-IQA QualiCLIP HPSv2.1 PickScore DPG Bench score GPT-4o preference win rate user study preference
Baselines: FLUX1.1 [Pro] Ultra PixArt-Sigma-XL Sana-1.6B FLUX-1.dev SDEdit I-Max Real-ESRGAN SinSR
Builds on: LoRA PiSSA classifier-free guidance guidance distillation SDEdit
Closest papers by meaning
- UltraPixel: Advancing Ultra High-Resolution Image Synthesis to New Peaks (NeurIPS 2024 · similarity 0.88)
- UltraHR-100K: Enhancing UHR Image Synthesis with A Large-Scale High-Quality Dataset (NeurIPS 2025 · similarity 0.88)
- ScaleCrafter: Tuning-free Higher-Resolution Visual Generation with Diffusion Models (ICLR 2024 · similarity 0.87)
- DiffuseHigh: Training-Free Progressive High-Resolution Image Synthesis Through Structure Guidance (AAAI 2025 · similarity 0.87)
- ResAdapter: Domain Consistent Resolution Adapter for Diffusion Models (AAAI 2025 · similarity 0.87)
- UDT: Reconciling U-Nets and Diffusion Transformers with Data-Adaptive Token Reduction (NeurIPS 2026 · 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.