Research map / Vision-language and multimodal / Visual question answering
FiLM: Visual Reasoning with a General Conditioning Layer
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use feature-wise affine modulation conditioned on the question | CNN+GRU visual reasoning model · CLEVR | accuracy: 95.5 -> 97.7 | Helped |
| Fine-tune the question-processing pipeline on CLEVR-Humans | CNN+GRU+FiLM · CLEVR-Humans | accuracy: 56.6 -> 75.9 | Helped |
| Add question-conditioned scaling and shifting together rather than fixing either parameter (ablation) | FiLM visual reasoning model · CLEVR validation | accuracy: 95.9 -> 97.4; 96.9 -> 97.4 | Helped |
| Allow gamma to take unrestricted values, including large, negative, or zero values (ablation) | FiLM visual reasoning model · CLEVR validation | accuracy: 95.9 -> 97.4; 96.3 -> 97.4 | Helped |
| Move FiLM within the residual block, including after ResBlock ReLU-2 (ablation) | FiLM visual reasoning model · CLEVR validation | accuracy: 96.6; 97.7; 97.1; 95.0 | No clear effect |
| Add FiLM conditioning in the ResBlocks (ablation) | FiLM visual reasoning model · CLEVR validation | accuracy: 21.4 -> 97.4 | Helped |
| Retain residual connections (ablation) | FiLM visual reasoning model · CLEVR validation | accuracy: 94.0 -> 97.4 | Helped |
| Retain batch normalization (ablation) | FiLM visual reasoning model · CLEVR validation | accuracy: 93.7 -> 97.4 | Helped |
| Use only 1x1 convolutions and omit coordinate maps (ablation) | FiLM visual reasoning model · CLEVR validation | accuracy: 97.4 -> 95.3 | Hurt |
| Use raw pixels instead of precomputed image features (ablation) | FiLM visual reasoning model · CLEVR validation | accuracy: 97.4 -> 97.6 | No clear effect |
| Reduce the number of ResBlocks to one (ablation) | FiLM visual reasoning model · CLEVR validation | accuracy: 97.4 -> 93.5 | Hurt |
| Use the zero-shot linear combination of FiLM parameters for unseen attribute combinations | FiLM model trained on CLEVR-CoGenT Condition A · CLEVR-CoGenT Condition B | accuracy: +3.2%; 71.5% -> 80.7% | Helped |
Closest papers by meaning
- MiCo: Multi-image Contrast for Reinforcement Visual Reasoning (NeurIPS 2025 · similarity 0.88)
- VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.87)
- SPARC: Separating Perception And Reasoning Circuits for Test-time Scaling of VLMs (ICML 2026 · similarity 0.87)
- Look, Remember and Reason: Grounded Reasoning in Videos with Language Models (ICLR 2024 · similarity 0.87)
- VGent: Visual Grounding via Modular Design for Disentangling Reasoning and Prediction (CVPR 2026 · similarity 0.86)
- SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · 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.