Upd. 06.10: V4 is out. 1) This is my gratitude to everyone supporting me and my work. I truly appreciate you guys! 2) Besides training the model further, I've also made some adjustments - the colors aren't THAT dark by default now compared to V3. 3) In my tests, V4 is the best of all the versions. But that's when it comes to my usage - it may be different for you. LUSTIFY! A photoreal checkpoint, capable of creating decent **** scenes without any loras. It understands both danbooru tags and natural language prompting, because it was trained like that. But that's not the only use - it can also create gorgeous SFW pics of objects/animals/fantasy creatures/you name it (you'll have to convince it tho - the model is horny). V1 - Good at ****, "meh" at everything else. V2 (recommended) - Good at aesthetic NSFW and SFW, sturggles with doggystyle, the rest is good. V3 - Combines V1's **** capabilites with V2's aesthetics. The visuals are darker than the other two versions (can be countered with prompting if you don't like the darker colors). A couple of things I need to make clear before you download the checkpoint: 1) All of my stuff is very opinionated. I create models mostly to suit my needs, my workflows, my prompting, etc. The preview images don't represent this, they use very basic prompts to show what the model can do with something really simple. 2) Sometimes you can get bad hands and bad faces at distant shots. These are hard to fix with finetuning, but really easy with inpainting/adetailer. So, please, don't bother me with the complaints about this. Either adapt to the needs of this checkpoint, or use something else (spoiler, 99% of checkpoints suffer from that because of the limitations of this tech). 3) Everything I do is experimental, and I do a lot of dumb-crazy stuff that in theory shouldn't work. So it's to be expected that something, what used to work with other checkpoint for you, might not work with this one. Let me know if there are such things, and I'd see if I can do anything about it. Recomended parameters: Sampler: DPM++ 2M SDE/DPM++ 3M SDE, Scheduler: Exponential/Karras Steps: 30 Cfg: 2.5-6. Highres.fix: upscale by 1.4-1.5, denoising ~0.4 For the distant shots of people you probably wanna use highres.fix and/or adetailer.
Also, use reasonable positive/negative prompts, shizoprompting does more harm than good (with any SDXL model). FAQ:
Q: "Output is broken is Fooocus". A: Turn off "Fooocus Enhance" style.
Q: "It can't do men". A: It can, check your prompt.
Some tags that have a big visual impact with this model:
Camera type tags: "shot on Polaroid SX-70", "shot on Kodak Funsaver", "shot on GoPro Hero", "shot on Canon EOS 5D", "shot on Leica T".
Photography styles: "analog photo", "glamour photography", "street fashion photography", "candid photo", "******* photo".
Lighting types: "cinematic lighting", "neon lighting", "soft lighting", "dramatic lighting", "low key lighting", "bright flash photography", "warm golden hour lighting", "radiant god rays".
Film types: "Ilford HP5 Plus", "Lomochrome color film", "Fujicolor Pro".
Photographers: Alessio Albi, Martin Schoeller, Miles Aldridge, Oleg Oprisco, Tim Walker.
Others: "film grain", "bokeh", "dreamy haze", "technicolor", "underexposed", "low quality", "lowres"
Alternatively, you can give a read to this wonderful prompting guide by PromptGeek. It has A LOT of useful tips I didn't know about even after 2 years of using SD. Work In progress: V1 started out as a mix of my personal unreleased SDXL finetune with such wonderful checkpoints as bigASP and Pyro's NSFW. Then I've finetuned this mix on a curated dataset of 1000 pics for aesthetic tweak. For V2 I've merged V1 with NatVis (20%) and then trained on new 1700 pics for additional 26 535 steps. V3 is a DARE merge of V1 and V2 that is then trained further. Why? Because I wanted to get V1's **** capabilities and V2's lighting+prompt adhearance.