Research map / Image generation / Fast diffusion sampling
LearniBridge: Learnable Calibration of Feature Caching for Diffusion Models Acceleration
Xuyue Huang, Zhe Chen, Wang Shen, Xiao-Ping Zhang
This paper tried a common change: Larger LoRA rank
increase LoRA rank beyond 64 · FLUX.1-dev · PSNR No clear effect
What 233 accepted papers report for this change
- 107 improved
- 60 no clear effect
- 19 mixed
- 47 hurt
- Understanding LoRA as Knowledge Memory: An Empirical Analysis ICML 2026 · increase LoRA rank · CounterFact · efficacyImproved
- SLQ: Bridging Modalities via Shared Latent Queries for Retrieval with Frozen MLLMs ICML 2026 · increase LoRA rank from 4 to 32 · COCO image-to-text · Recall@5: 84.5 -> 82.1Hurt
- Trojan-Speak: Bypassing Constitutional Classifiers with No Jailbreak Tax via Adversarial Finetuning ICML 2026 · Increase LoRA rank (8 to 128) · GPQA-Diamond during Stage 2 · accuracyImproved
See this change in detailMatched by the wording of each paper's own result, read by AI from the full text. Papers report gains more often than failures, so this shows what gets published, not the chance a change works.
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 LoRA calibration of the final block to cached features (LearniBridge vs TaylorSeer at N=5) | feature caching on FLUX.1-dev · DrawBench | ImageReward: 0.9359 -> 0.9590 | Improved |
| add LoRA calibration (LearniBridge vs TaylorSeer at N=5) on HunyuanVideo | feature caching on HunyuanVideo · VBench | VBench (%): 80.13 -> 79.93 | Hurt |
| add LoRA calibration (LearniBridge vs TeaCache at similar speed) on WAN 2.1 | feature caching on WAN 2.1-1.3B · VBench | VBench (%): 80.04 -> 81.21 | Improved |
| increase number of calibration prompts from 1 to 5 (ablation) | LearniBridge · FLUX.1-dev calibration prompts | PSNR: 27.7913 -> 30.1623 | Improved |
| increase number of calibration prompts beyond 5 (ablation) | LearniBridge · FLUX.1-dev calibration prompts | PSNR: 30.1623 -> 30.0814 (30 prompts) | No clear effect |
| use medium or long calibration prompts instead of short ones (ablation) | LearniBridge · FLUX.1-dev calibration prompts | SSIM: 0.6927 -> 0.7898 | Improved |
| increase LoRA rank beyond 64 (ablation) | LearniBridge · FLUX.1-dev | PSNR | No clear effect |
| add LoRA adapters on output projection and feed-forward layers (ablation) | LearniBridge · FLUX.1-dev | PSNR | Improved |
| add LoRA adapters on query, key, value projections (ablation) | LearniBridge · FLUX.1-dev | PSNR | Improved |
Closest papers by meaning
- Forecast Then Calibrate: Feature Caching as ODE for Efficient Diffusion Transformers (AAAI 2026 · similarity 0.93)
- HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 2025 · similarity 0.92)
- Beyond Fixed Formulas: Data-Driven Linear Predictor for Efficient Diffusion Models (CVPR 2026 · similarity 0.91)
- ProCache: Constraint-Aware Feature Caching with Selective Computation for Diffusion Transformer Acceleration (AAAI 2026 · similarity 0.91)
- Accelerating Diffusion Transformers with Token-wise Feature Caching (ICLR 2025 · similarity 0.91)
- FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality (ICLR 2025 · similarity 0.90)
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.
