The authors introduce visual abstention, the ability to identify that no valid image transformation exists, explain the conflict and decline to generate.
Multimodal models can learn when impossible image edits require refusal
Paired feasibility training taught one model to reject impossible transformations while improving its accuracy on valid edits.
Academic
Chufan Shi · Cheng Yang · Tiannuo Yang · Isadora White · Yiwei Chen · Taylor Berg-Kirkpatrick · +1 more
University of Southern California · University of California San Diego
Research Digest··3 min read
Shi and colleagues test whether unified multimodal models can recognize when an image-editing request is impossible under the stated constraints, rather than generating a misleading approximation.
Why this paper
From University of Southern California and University of California San Diego
In one line
Unified multimodal models can be trained to refuse impossible image edits, achieving 93% refusal without sacrificing editing accuracy.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§