Research map / Label-efficient and robust learning / Test-time adaptation
BiProLoRA: Bilevel Prompt LoRA for Real Scene Recovery
Nan An, Long Ma, Tengyu Ma, Zhu Liu, Yingchi Liu, Risheng Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add self-supervised distribution-fidelity learning (DFL) before joint modeling (ablation) | BiProLoRA · DARKFACE | NIQE: 3.403 -> 2.971 | Improved |
| add DFL (ablation) | BiProLoRA · DARKFACE | DE: 7.180 -> 7.201 | Mixed |
| finetune pretrained weights during DFL instead of adapters only (ablation) | BiProLoRA with DFL · DARKFACE | NIQE: 2.971 -> 3.607 | Hurt |
| add prompt embeddings on top of LoRA through bilevel HO (ablation) | BiProLoRA with DFL · DARKFACE | NIQE: 4.074 -> 2.971 | Improved |
| use prompt embeddings only without LoRA (ablation) | BiProLoRA with DFL · DARKFACE | NIQE: 4.550 vs 4.074 for LoRA only | Hurt |
| replace naive combination of synthetic and real objectives with bilevel HO (ablation) | BiProLoRA with DFL · DARKFACE | NIQE: 3.716 -> 2.971 | Improved |
| train DFL on synthetic LOL images instead of real degraded data (ablation) | DFL · DARKFACE | texture fidelity (qualitative) | Hurt |
| train DFL on different real degradation sets (RTTS, URPC, DARKFACE) (ablation) | DFL · DARKFACE | texture fidelity (qualitative) | No clear effect |
| lower penalty coefficient lambda to 0.1 (ablation) | BiProLoRA · DARKFACE | LIQE, DE, MUSIQ, ARNIQ | Hurt |
| raise penalty coefficient lambda to 0.3 (ablation) | BiProLoRA · DARKFACE | LIQE, DE, MUSIQ, ARNIQ | Hurt |
| preprocess low-light images with BiProLoRA before detection | Cascade R-CNN · ExDark nighttime detection | mAP: 37.5 -> 40.9 | Improved |
| preprocess with Di-Retinex, LightenDiff or ReDDiT before detection | Cascade R-CNN · ExDark nighttime detection | mAP: 37.5 -> 36.2 (Di-Retinex) | Hurt |
Closest papers by meaning
- ReFIR: Grounding Large Restoration Models with Retrieval Augmentation (NeurIPS 2024 · similarity 0.85)
- Compositional Scene Understanding through Inverse Generative Modeling (ICML 2025 · similarity 0.84)
- UP-Restorer: When Unrolling Meets Prompts for Unified Image Restoration (AAAI 2025 · similarity 0.84)
- Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild (CVPR 2024 · similarity 0.84)
- LL-Bench: Rethinking Low-Level Vision Evaluation in the Era of Large-Scale Generative Models (NeurIPS 2026 · similarity 0.84)
- Cached Multi-Lora Composition for Multi-Concept Image Generation (ICLR 2025 · similarity 0.84)
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.
