Had some time to play recently Trying to get as much detail as possible, unfortunately the further out the shot the more the face gets mangled, but I'm happy with these, enjoy
messg said: Great work on the lora. I'm presuming it's SDXL made.
Thanks, it is SDXL but its actually a fine tuned model, these are direct txt2img with an upscale which tbh I didnt realise means they get displayed full size here so I'll not bother upscaling in future.
BigmacTom said: This model seems to do really well with faces.... how does it cope with bodies and clothes?
Maybe if I spend some time working with it a bit more I might be able to do better but with simple prompts and no exploration of settings the wider shots are not at the level I like.
Added some examples but I didn't spend much time on them.
deadpool said: I thought that I'd give it a try.
Close up images of messy Latinas. Some are in the midst of getting messy; some have already gone through it.
These are from Bing, if you're curious.
Yeah I cant get Dalle to anything, it might do men like that for me but def not women (using copilot), I keep telling it that its sexist and it closes the conversation lol
My issue with Dalle is the lack of realism and the massive mouths it usually does bu thats my preference, I think it really excels at understanding complex scenes.
messg said: Great work on the lora. I'm presuming it's SDXL made.
Thanks, it is SDXL but its actually a fine tuned model, these are direct txt2img with an upscale which tbh I didnt realise means they get displayed full size here so I'll not bother upscaling in future.
When you say finetune, are you talking about a full SDXL finetune or a Lora? A full wam/messy finetune would be immense but the amount of images to sort and caption and then the resources to process it have killed that project for me. I've had some good success with LORA's but with SD3 due soon, I've held off working on any more until it drops.
It's actually a merge of 6 Lora at various weights into an existing model, so there's no Lora anymore and just a model but it's not a new model either. Merge is probably the best definition.
kemistry said: It's actually a merge of 6 Lora at various weights into an existing model, so there's no Lora anymore and just a model but it's not a new model either. Merge is probably the best definition.
A decent approach as you'll be able to fine tune each lora for individual specifics before merging them rather than trying to do everything at once.
I looked at fine tuning an existing model like Juggernaut which should be manageable on a 4090 albeit slow but Merging Loras seems a more logical approach as you can dial in each lora before the merge. Either way, with SD3 due soon and it's captioning being more developed I put the work on pause until it releases. With Dalle3/Bing becoming more and more restricted, SD is definitely the way forward.
kemistry said: Had some time to play recently Trying to get as much detail as possible, unfortunately the further out the shot the more the face gets mangled, but I'm happy with these, enjoy
Close up images of messy Latinas. Some are in the midst of getting messy; some have already gone through it.
These are from Bing, if you're curious.
could you share the prompt for these?
Certainly.
color photo by walter iooss jr of a 2010s tv scene featuring a spunky 25yo Latina in a black shirt. close up of her face. a torrent of VERY LUMPY green OATMEAL MOSTLY covers her face. it pours out from high above in a stream, splattering w/ motion blur. the same substance slimes her shirt. she has an embarrassed smile. white background low light film grain. the scene shows a fun mood
(the "torrent" part doesn't have to be there; the oatmeal can still be splattered on her face, though)
kemistry said: Had some time to play recently Trying to get as much detail as possible, unfortunately the further out the shot the more the face gets mangled, but I'm happy with these, enjoy