Use case · Idea to test
Stop an unsafe image before it finishes
Check during generation and cancel early.
- Image generation in progress
- Classifier checks a checkpoint
- Cancel or keep generating
An image service could cancel a generation once an intermediate classifier detects a policy violation, rather than paying for every remaining generation step and the final image decode.
Where the small model fits
A trained safety head could read intermediate image-model features, or classify a cheap partial preview. The first route might avoid decoding a preview; the second adds some decoding work. In a diffusion model, checkpoints would be denoising steps, not necessarily the transformer layers used in our text experiments.
What would need to work
Finish matched generations offline, using the same prompts and seeds, to check what cancelled images would have become. Measure unsafe outputs missed, acceptable images cancelled, useful checkpoint depth and net time after preview/checking costs. Passing an early check does not certify the final image; continuing outputs still need the final check. This experiment has not been run here.