V0.2 was trained on a completely overhauled training set to reduce biases towards specific faces, outfits, and poses:<ul><li>No repeating backgrounds and faces (I inpainted over every single face and any distracting backgrounds in the training set)</li><li>More diverse outfits (50% or more of the subjects were wearing denim jeans in the previous version)</li><li>More diverse poses (spread toes, feet touching/apart, feet up, knees flexed) but can be hard to control in generations</li><li>Includes clean and dirty feet (specify preference in prompt)</li></ul>Tip: If you like a generated image but it has missing/extra toes, try doing variations of the image with the same prompt at moderate strength (0.5-0.6) and it will usually fix it after a few tries. It's much faster and usually more coherent than inpainting if you're not too attached to any other minor details.Training was based on the checkpoint <a target="_blank" rel="ugc" href="https://civarchive.com/models/238116/opendalle">OpenDalle v1.1</a>, but it seems to work reasonably well with other checkpoints. It does minor variations of a basic barefoot seated pose on the ground with the soles of the feet facing the viewer that every model I tested so far was unable to produce coherently.

Seated pose, focus on soles of feet

Completely overhauled training set to reduce biases towards certain faces and outfits:<ul><li>No repeating backgrounds and faces (inpainted over every face and several backgrounds in the training set)</li><li>More diverse outfits (50% or more of the subjects were wearing denim jeans in the previous version)</li><li>More diverse poses (spread toes, feet touching/apart, feet up, knees flexed) but can be hard to control in generations</li><li>Includes clean and dirty feet (specify preference in prompt)</li></ul>

Files with SHA256: a6eca2fe09458d6bb71fff3be13840e3f58b8b5d050da333d8811239535cf6e7