Baseline corpus: ~10K generated images and LlavaGuard-7B safety annotations from 7 existing publicly-released PRX checkpoints. Provides comparison points for the controlled dose-response variants in anonym371/dose-response-generated-images.

Repository structure

images/
├── prx-1024-beta/images.tar
├── prx-256-base/images.tar
├── prx-256-sft/images.tar
├── prx-512-base/images.tar
├── prx-512-dc-ae/images.tar
├── prx-512-sft/images.tar
└── prx-512-sft-distilled/images.tar

annotations/
├── prx-1024-beta.parquet
├── prx-256-base.parquet
├── prx-256-sft.parquet
├── prx-512-base.parquet
├── prx-512-dc-ae.parquet
├── prx-512-sft.parquet
└── prx-512-sft-distilled.parquet

Annotation schema

Column	Type	Description
`id`	string	Image filename stem (matches an entry in the paired `images.tar`)
`rating`	string	`Safe` or `Unsafe`
`category`	string	One of O1–O9 (when unsafe) or `NA: None applying` (when safe)
`rationale`	string	LlavaGuard's free-text reasoning

Safety judge

AIML-TUDA/LlavaGuard-v1.2-7B-OV — vision-language safety classifier.

Plus NA: None applying for safe outputs.

Source models

The 7 PRX checkpoints are the public ones released by Photoroom:

Photoroom/prx-1024-t2i-beta (and the other PRX variants documented in the PRX framework)

Images here are generated by these public PRX checkpoints — distinct from the dose-response-trained variants in anonym371/dose-response-generated-images.