Research map / Image generation / Fast diffusion sampling
LINA: Learning INterventions Adaptively for Physical Alignment and Counterfactual Generation in Diffusion Models
Shu Yu, Chaochao Lu
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 token-level intervention γ1 (ablation) | Lina (SD-3.5) · PAP-Optics | Success rate (%): 85.1 -> 97.4 | Improved |
| add token-level intervention γ1 (ablation) | Lina (SD-3.5) · Winoground | Success rate (%): 60.2 -> 79.5 | Improved |
| add latent-level contrastive guidance γ2 (ablation) | Lina (SD-3.5) · PAP-Optics | Success rate (%): 81.3 -> 97.4 | Improved |
| add latent-level contrastive guidance γ2 (ablation) | Lina (SD-3.5) · PAP-Density | Success rate (%): 78.0 -> 92.3 | Improved |
| add adaptive intervention module instead of fixed γ (ablation) | Lina (SD-3.5) · Winoground | Success rate (%): 68.4 -> 79.5 | Improved |
| add adaptive intervention module instead of fixed γ (ablation) | Lina (SD-3.5) · PAP-Density | Success rate (%): 85.2 -> 92.3 | Improved |
| use reallocated causality-aware schedule instead of standard (ablation) | Lina (SD-3.5) · PAP-Optics | Success rate (%): 92.3 -> 97.4 | Improved |
| use reallocated causality-aware schedule instead of standard (ablation) | Lina (SD-3.5) · Winoground | Success rate (%): 74.3 -> 79.5 | Improved |
| apply Lina to SD-3.5-large | SD-3.5-large · PAP-Density | Success rate (%): 54.2 -> 92.3 | Improved |
| apply Lina to FLUX.1 | FLUX.1-Krea-dev · Winoground | Success rate (%): 65.5 -> 83.0 | Improved |
| add LoRA finetuning to SD-3.5 | SD-3.5-large · Winoground | Success rate (%): 54.4 -> 57.3 | No clear effect |
| add PPAD to SD-3.5 | SD-3.5-large · Winoground | Success rate (%): 54.4 -> 62.6 | No clear effect |
Closest papers by meaning
- Stitched Value Model for Diffusion Alignment (NeurIPS 2026 · similarity 0.88)
- Learning Diffusion Models with Flexible Representation Guidance (NeurIPS 2025 · similarity 0.87)
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.87)
- REPA Works Until It Doesn’t: Early-Stopped, Holistic Alignment Supercharges Diffusion Training (NeurIPS 2025 · similarity 0.87)
- Composition and Alignment of Diffusion Models using Constrained Learning (NeurIPS 2025 · similarity 0.87)
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 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.
