Research map / Image generation / Image restoration
HAODiff: Human-Aware One-Step Diffusion via Dual-Prompt 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 HAODiff, a one-step diffusion model for human body restoration that adds human motion blur to the degradation pipeline and uses a triple-branch dual-prompt guidance (DPG) to produce positive/negative prompts for classifier-free guidance.
Problem
Restoring human-centric images that suffer from both generic degradation and human motion blur (HMB), which existing degradation pipelines and restoration models under-cover.
Approach
A degradation pipeline that synthesizes HMB using Sapiens body-part masks blended with globally blurred images, and a DPG module whose three branches predict HQ image, residual noise (LQ minus HQ), and HMB mask to form adaptive positive and negative prompts for CFG in a single diffusion step.
Key results
- On synthetic PERSONA-Val, HAODiff reaches DISTS 0.1023 and FID 8.3623, best among compared methods, with inference 0.20 s per 512x512 image on RTX A6000.
- On MPII-Test, HAODiff achieves the lowest HMB-R of 0.0929 versus 0.1853 for the next best (OSEDiff), and tops all listed no-reference metrics.
- Replacing DPG with fixed text-pair CFG, DAPE, or HFIE prompts yields worse DISTS, FID, and MS-SWD, indicating DPG reduces color shift.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the HMB-aware third DPG branch (ablation) | HAODiff · MPII-Test | HMB restoration quality | Helped |
| Use DPG instead of fixed positive-negative prompts, DAPE, or HFIE | HAODiff · PERSONA-Val | restoration quality | Helped |
| Use DPG instead of fixed positive-negative prompts, DAPE, or HFIE | HAODiff · PERSONA-Val | color distribution difference (MS-SWD) | Helped |
Surprising findings
- Competing prompt-based methods produce noticeable color shifts that are hard to remove with post-processing, which DPG's negative residual-noise prompt reduces.
Limitations
- Heavy global degradation combined with strong local motion blur on fast-moving limbs remains challenging; limb poses can be unnatural and overlapping objects may be misrestored.
- CFG requires two UNet passes, giving a modest efficiency penalty relative to OSEDiff and OSDHuman.
Experimental setup
Datasets: PERSONA LSDIR FFHQ MPII Human Pose PERSONA-Val PERSONA-Test
Benchmarks: PERSONA-Val PERSONA-Test MPII-Test
Metrics: DISTS LPIPS TOPIQ FID CLIPIQA MANIQA NIQE LIQE IL-NIQE HMB-R
Baselines: SUPIR DiffBIR SeeSR PASD ResShift SinSR OSEDiff InvSR OSDHuman S3Diff text pair DAPE HFIE
Builds on: OSDHuman OSEDiff Real-ESRGAN SwinIR classifier-free guidance
Closest papers by meaning
- One-Step Effective Diffusion Network for Real-World Image Super-Resolution (NeurIPS 2024 · similarity 0.86)
- BlurDM: A Blur Diffusion Model for Image Deblurring (NeurIPS 2025 · similarity 0.86)
- Identifying and Solving Conditional Image Leakage in Image-to-Video Diffusion Model (NeurIPS 2024 · similarity 0.86)
- EVODiff: Entropy-aware Variance Optimized Diffusion Inference (NeurIPS 2025 · similarity 0.86)
- Single Shot HDR Recovery via a Video Diffusion Prior (NeurIPS 2026 · similarity 0.86)
- DEFT: Efficient Fine-tuning of Diffusion Models by Learning the Generalised $h$-transform (NeurIPS 2024 · 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.