Update: Following a request, i reposted the old versions "IXL Next" and "IXL V0" here as previous versions of Reborn.Usage<ul><li>Steps 25-30 should be enough, 40 doesn't hurts :D</li><li>CFG: i tested with 5, going lower may be fine</li><li>Scheduler Sampler: Euler A and DPM++ 2M should be fine</li><li>Clip Skip: the usual 2</li><li>Adetailer: recommended :D</li><li>Negative prompt: IllusN3g, deep_negative_pony (for a more authentic feeling :D)</li></ul>Adetailer trick: To avoid the "box effect", what i usually do is set this in the inpaint section:<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/b00e5c88-2006-4b2a-8b1b-999e592e0a4e/width=525/b00e5c88-2006-4b2a-8b1b-999e592e0a4e.jpeg" />Long Story™One thing has been frustrating me for a long time. Neither the Next or V0 version of BancinIXL felt close enough to BancinXL. So, i deleted those got back to work.Here, i started back from scratch to try and get a closer feeling, using two different approach using AnBan 1.0:<ul><li>Using supermerger with "Add Difference" in "Train difference" using BancinXL-Pony (leveraging xyz plot, i got that 0.5 was the max i could incorporate this way)</li><li>Directly finetuning the checkpoint using the dataset (200 pictures) i previously used for a BancinXL LoRA (the method was the same used for HoJ, but since the dataset was small, it got a strong bias)</li></ul>This left me with two differents checkpoints that i merged together with an intermediate step for each (closer to merge, closer to finetune), carefully selecting the block pairs that i liked.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/b988016c-6f1c-4b23-8317-f792f21b174d/width=525/b988016c-6f1c-4b23-8317-f792f21b174d.jpeg" />Feel free to experiment, this is still not a "Pony like illustrious", but it should feel more western/comic style that AlterBan or Quirk 😅

Files with SHA256: bc15d2c1a35f42a1bb7b62b4c9eb43df7174a8422dd7d0a641400de26e380080