This was an experiment i did some time ago for this bounty: <a target="_blank" rel="ugc" href="https://civarchive.com/bounties/136/luisaplora-challenge-follow-my-settings-sdxl-10">https://civarchive.com/bounties/136/luisaplora-challenge-follow-my-settings-sdxl-10</a> It was asked to make a lora using a very limited set of pictures (only 5). Unluckly, the instructions given didn't work well for me, and i wasn't satisfied at all with the result. I then decided to try again much later, trying a different approach. seems like i did it. You can find here the lora, the dataset used for training, and the json containing the training settings for koya_ss gui. It was trained on a RX5700XT (8gb vram), on Linux. On Nvidia cards you can use 8-bit AdamW and xformers instead of sdpa. trigger for the model is drizzt_dnd

Files with SHA256: 441947881acdd3e92b79d04f3a17567e97fc9c780e28c750fdf29b06e1d85770