Post
32
Let me introduce you to our CVPR 2026 paper!
Today's content moderation systems give you a label: safe or unsafe. They don't tell you what triggered the decision, who is involved, or where in the image it happens. That opacity hurts auditing, breaks adaptation across platforms, and frustrates the human review that responsible deployment demands.
We built SenBen to fix this: the first large-scale scene graph benchmark designed specifically for sensitive content moderation:
- 13,999 annotated frames from 157 movies
- Visual Genome style scene graphs with bounding boxes, attributes, and predicates
- Affective state attributes (pain, fear, aggression, distress) so the model captures not just what is in the frame, but what it means
- 16 safety tags across 5 categories, the broadest taxonomy of any dataset of this kind
A small model that beats much bigger ones:
We distilled a frontier VLM into a compact 241M parameter student built on Florence-2.
On grounded scene graph metrics, the 241M student beats every evaluated VLM except Gemini, and every commercial safety API. It also wins on object detection and captioning across the entire model zoo. It runs at 733 ms per frame on 1.2 GB VRAM, which is 7.6 times faster than the next-best local VLM at zero per-frame cost. The whole benchmark, from dataset creation through all baseline evaluations, is reproducible for under $350.
Project: https://senben.kim/
Paper: SenBen: Sensitive Scene Graphs for Explainable Content Moderation (2604.08819)
Dataset: fcakyon/senben
Code (soon): https://github.com/fcakyon/senben
Today's content moderation systems give you a label: safe or unsafe. They don't tell you what triggered the decision, who is involved, or where in the image it happens. That opacity hurts auditing, breaks adaptation across platforms, and frustrates the human review that responsible deployment demands.
We built SenBen to fix this: the first large-scale scene graph benchmark designed specifically for sensitive content moderation:
- 13,999 annotated frames from 157 movies
- Visual Genome style scene graphs with bounding boxes, attributes, and predicates
- Affective state attributes (pain, fear, aggression, distress) so the model captures not just what is in the frame, but what it means
- 16 safety tags across 5 categories, the broadest taxonomy of any dataset of this kind
A small model that beats much bigger ones:
We distilled a frontier VLM into a compact 241M parameter student built on Florence-2.
On grounded scene graph metrics, the 241M student beats every evaluated VLM except Gemini, and every commercial safety API. It also wins on object detection and captioning across the entire model zoo. It runs at 733 ms per frame on 1.2 GB VRAM, which is 7.6 times faster than the next-best local VLM at zero per-frame cost. The whole benchmark, from dataset creation through all baseline evaluations, is reproducible for under $350.
Project: https://senben.kim/
Paper: SenBen: Sensitive Scene Graphs for Explainable Content Moderation (2604.08819)
Dataset: fcakyon/senben
Code (soon): https://github.com/fcakyon/senben
