Reward
Diversity
Validation samples over training
⚠️ Content warning. These are raw model samples from RL post-training. Pretrained visual generative models are usually audited to suppress violent or pornographic outputs, but visual post-training and diversity-promoting optimization can weaken those built-in safeguards and reintroduce harmful visual content.
step
1000
drag to scrub through training
One row per run: 3 samples of the same prompt at the selected step, using the step values logged in SwanLab. Rows fade out once a run has ended. Click an image to enlarge.
⚠️
Content notice
The samples below are unfiltered model outputs from RL post-training and may contain violent, sexual or otherwise disturbing imagery.