Th point of this checkpoint is to make it feel like an illustration and look almost like a photo. A painting behind a mist. It works for some themes quite well, I reckon. It may depend on the prompt.It is based on Rogue Pony merge, finetuned with several lora passes. So I'd take the checkpoint, train a lora using a set of 200-300 pictures, see which weight works best, merge lora to checkpoint and train another set on the result. Training sets were rendered in SDXL models and other Pony checkpoints, selected and tagged (semi) manually. I'd use parts of input prompt, excluding any quality tags. Parse it in Deepbooru captioning and correct the result, basically delete all the nonsense, which would be about the half of the output. Then possibly add some tags I thought are appropriate.I'd use DPM++ 2M SDE scheduler, cfg 4-5, about 30 steps. Optionally use score tags and some quality tags like "photorealistic" to push it more toward photo-like. Or reduce steps, increase cfg and add illustration related tags to push more toward artwork feel.You may notice I barely use negative tags for sample images. I think it works better that way. Have fun.

Files with SHA256: 8f52c0f5ed04cd9a52169198cc9f9034fd74edff114f7d2f5d984c08a69e9387