Research map / Image generation / Text-to-image diffusion
HiFlow: Training-free High-Resolution Image Generation with Flow-Aligned Guidance
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 HiFlow, a training-free, model-agnostic method that builds a virtual high-resolution reference flow from the low-resolution sampling trajectory to guide high-resolution flow-model generation via initialization, direction, and acceleration alignment.
Problem
Pre-trained text-to-image flow and diffusion models degrade in quality and structure when generating high-resolution images (2K-4K); existing training-free methods rely only on the final low-resolution endpoint, causing artifacts and weak detail fidelity.
Approach
A virtual reference flow in high-resolution space is built from step-wise predicted clean samples of the low-resolution trajectory; it guides high-res sampling by (1) initializing from a noised reference state, (2) low-pass Fourier alignment of the denoising direction, and (3) aligning the flow's acceleration (change of velocity) with the reference's.
Key results
- At 2K, HiFlow reaches FID 55.39 and FID_patch 47.70, the best among compared methods (DemoFusion 56.07 FID as second best).
- At 4K, HiFlow reaches FID 52.55 and FID_patch 45.01, beating I-Max (53.27) and UltraPixel (47.19 patch FID second best).
- HiFlow runs at 56 s for 2K and 379 s for 4K, faster than DemoFusion, DiffuseHigh and I-Max on a single A100.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add initialization, direction, and acceleration alignment (ablation) | HiFlow · 2K and 4K | FID: 67.87 -> 55.39; 234.38 -> 52.55 | Helped |
| Add direction alignment (ablation) | HiFlow without acceleration alignment · 2K and 4K | FID: 58.40 -> 55.39; 55.36 -> 52.55 | Helped |
| Add acceleration and direction alignment (ablation) | HiFlow without acceleration or direction alignment · 2K and 4K | FID: 58.26 -> 55.39; 56.20 -> 52.55 | Helped |
| Add acceleration alignment (ablation) | HiFlow · 2K and 4K | FID_patch: 50.38 -> 47.70; 51.79 -> 45.01 | Helped |
| Use HiFlow instead of DemoFusion | Flux.1.0-dev high-resolution generation · 2K | latency (sec.): 106 -> 56 | Helped |
| Use HiFlow instead of DiffuseHigh | Flux.1.0-dev high-resolution generation · 4K | latency (sec.): 533 -> 379 | Helped |
Surprising findings
- Ablation shows removing acceleration alignment alone yields a higher CLIP score (35.44 vs 35.40 at 4K) and higher IS at 2K, indicating a small trade-off against the full model.
- Diffusion-4k, a training-based method, performs worst on 4K FID (77.95) among the compared methods.
Limitations
- Generation quality depends on the quality of the low-resolution reference flow; structural anomalies in the reference (e.g., duplicated tiger tail) are propagated.
- Evaluated mainly on Flux.1.0-dev with a 1K-caption prompt set; FID reference uses LAION-High-Resolution, so absolute FID values are modest in differences.
Experimental setup
Datasets: LAION-High-Resolution
Benchmarks: LAION-High-Resolution (FID reference)
Metrics: FID FID_patch IS IS_patch CLIP score latency
Baselines: DemoFusion DiffuseHigh I-Max Flux + BSRGAN UltraPixel Diffusion-4k
Builds on: Rectified Flow DemoFusion I-Max DiffuseHigh
Closest papers by meaning
- Boosting Flow-based Generative Super-Resolution Models via Learned Prior (CVPR 2024 · similarity 0.88)
- Scaling Rectified Flow Transformers for High-Resolution Image Synthesis (ICML 2024 · similarity 0.88)
- DiffuseHigh: Training-Free Progressive High-Resolution Image Synthesis Through Structure Guidance (AAAI 2025 · similarity 0.88)
- ResMaster: Mastering High-Resolution Image Generation via Structural and Fine-Grained Guidance (AAAI 2025 · similarity 0.87)
- Align Your Flow: Scaling Continuous-Time Flow Map Distillation (NeurIPS 2025 · similarity 0.87)
- STARFlow: Scaling Latent Normalizing Flows for High-resolution Image Synthesis (NeurIPS 2025 · 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.