Research map / Vision-language and multimodal / Human and embodied modeling
LANCE: Stress-testing Visual Models by Generating Language-guided Counterfactual Images
Viraj Uday Prabhu, Sriram Yenamandra, Prithvijit Chattopadhyay, Judy Hoffman
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 |
|---|---|---|---|
| structured language-guided counterfactual edits vs original images | ResNet-50 · Hard ImageNet | acc@1: 79.86 -> 74.01 | Hurt |
| LANCE counterfactuals on ViT-B | ViT-B · Hard ImageNet | acc@1: 85.42 -> 79.00 | Hurt |
| LANCE counterfactuals on ConvNext | ConvNext · Hard ImageNet | acc@1: 85.42 -> 81.18 | Hurt |
| LANCE counterfactuals on CLIP | CLIP · Hard ImageNet | acc@1: 56.25 -> 48.27 | Hurt |
| LANCE structured perturbation vs random perturbation LANCE-R (ablation) | ResNet-50 · Hard ImageNet | acc@1: 78.81 -> 74.01 | Hurt |
| LANCE vs LANCE-R on CLIP (ablation) | CLIP · Hard ImageNet | acc@1: 50.13 -> 48.27 | Hurt |
| LANCE vs LANCE-R on ConvNext (ablation) | ConvNext · Hard ImageNet | acc@1: 85.79 -> 81.18 | Hurt |
| image reconstruction without edits (control) (ablation) | ResNet-50 · Hard ImageNet | acc@1: 79.86 -> 79.86 | No clear effect |
| reconstruction control on CLIP (ablation) | CLIP · Hard ImageNet | acc@1: 56.25 -> 55.90 | No clear effect |
| reconstruction control on ConvNext (ablation) | ConvNext · Hard ImageNet | acc@5: 98.26 -> 98.09 | No clear effect |
| LANCE on ResNet-50 acc@5 | ResNet-50 · Hard ImageNet | acc@5: 95.83 -> 92.96 | Hurt |
| background edits vs other edit types (ablation) | pretrained models · Hard ImageNet | relative accuracy change | Hurt |
Closest papers by meaning
- Free on the Fly: Enhancing Flexibility in Test-Time Adaptation with Online EM (CVPR 2025 · similarity 0.84)
- VIGC: Visual Instruction Generation and Correction (AAAI 2024 · similarity 0.84)
- CogCoM: A Visual Language Model with Chain-of-Manipulations Reasoning (ICLR 2025 · similarity 0.84)
- ImagenHub: Standardizing the evaluation of conditional image generation models (ICLR 2024 · similarity 0.84)
- Diversity Matters: Revisiting Test-Time Compute in Vision-Language Models (ICML 2026 · similarity 0.84)
- Benchmarking Counterfactual Image Generation (NeurIPS 2024 · 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.
