InGuard moves safety checks inside the text-to-image generation pipeline, using the model's own text embeddings and latent states rather than external classifiers. It grades prompts as unsafe, risky, or benign, modifies risky prompts to produce safe images, and can stop generation early when risk is detected mid-denosing. This achieves 97.9-98.8% safety rate across five open-weight models, with less benign disturbance and significant speedups, while leaving base-model parameters untouched.
Inner pipeline safety checks match outer guardrails with fewer false positives
InGuard embeds risk scoring, embedding repair and a mid-denoising latent detector inside text-to-image models, cutting benign disturbance and skipped denoising steps across five open-weight models.
Chinese Tech
Zeyu Wang · Xiaodan Li · Zhiwen Li · Yuefeng Chen · Hui Xue
Alibaba AAIG
Research Digest··1 min read
In one line
Safety checks inside the T2I pipeline, using the model's own embeddings and latent states, can match or exceed outer guardrails with less disruption and cost.
Before you read this
Three ideas this paper builds onsafety alignmentProcess of training models to recognize and refuse harmful requests, ensuring outputs remain safe.116 papers →diffusion modelA type of generative model that learns to reverse a gradual noising process to create data.197 papers →activation steeringActivation steering modifies a model's internal vector representations during inference to shift behavior, such as editing outputs, removing capabilities, or controlling trust.83 papers →
Read it as
Wang et al. from Alibaba AAIG propose InGuard, a safety framework that operates inside the T2I pipeline using the model's own text encoder embeddings and intermediate latent states.
The authors designed a three-component inner guardrail for text-to-image (T2I) generation.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.