<p>After seeing the new detailed prompting methods flux uses to generate images, I decided to see if it was possible to train other models to accept JoyCaption style prompts. I will attempt to train multiple common non-Flux checkpoints like Pony, SDXL, and SD15 to accept more advanced prompts. Check each version's version details to see what it does and how it should be used.</p>

JoyCaption style prompting for non flux models

<p>Very rough beta version for SD1.5 trained on about 20 images at 512X512 res on kohayaSS at a relatively high strength for about 1500 steps. Seems to be a bit overtrained, so I recommend lower strengths of around 0.4 to 0.6, but you can go higher if you are upscaling. Results at 712 res seem to be good, and I recommend using DPM++ 2M SDE Karras as the sampler unless you are upscaling. Sadly doesnt seem to be any better at generating text than default SD1.5, but its prompt adherence to detailed prompts seems a bit better.</p>

Files with SHA256: 5e1c7b204f9dd74eff14a9837f3d85b93f52d63d69aa3912a9e5f62f7c9a7e76