Safety stacks beyond the loss function

Minimizing denoising loss does not prevent misuse. Production stacks layer refusal policies, classifiers on prompts and outputs, watermarking research, and dataset governance [2]. Static filters are incomplete as adversarial misuse evolves [2].

Media safety differs from text safety: visual classifiers operate on pixels or latents, and false positives can block benign artistic content if thresholds are tuned without domain review [2].

Prompt injection against tool-augmented systems tries to override policies via crafted user or tool text [2]. Content classifiers flag toxic or policy-violating media before display. Datasheets record provenance, consent limits, and known biases [2].

User-facing controls (negative prompts, seeds, strength sliders, ControlNet-style conditioning, inpainting masks) shape outputs without retraining base weights [2].

Safety for media includes provenance metadata and output classifiers for NSFW or violence. None replace human review for high-stakes publishing workflows [2].

Watermark detectors remain probabilistic; combine statistical tests with policy on how to handle ambiguous outputs in user-facing products [2].

Age-gating, geofencing, and terms-of-use enforcement sit outside the model weights but shape who can access which capabilities in consumer apps [2].

Incident response playbooks for generative media should cover takedown workflows when classifiers miss policy-violating outputs [2].