V0.2 was trained on a completely overhauled training set to reduce biases towards specific faces, outfits, and poses:
No repeating backgrounds and faces (I inpainted over every single face and any distracting backgrounds in the training set)
More diverse outfits (50% or more of the subjects were wearing denim jeans in the previous version)
More diverse poses (spread toes, feet touching/apart, feet up, knees flexed) but can be hard to control in generations
Includes clean and dirty feet (specify preference in prompt)
Tip: If you like a generated image but it has missing/extra toes, try doing variations of the image with the same prompt at moderate strength (0.5-0.6) and it will usually fix it after a few tries. It's much faster and usually more coherent than inpainting if you're not too attached to any other minor details.
Training was based on the checkpoint OpenDalle v1.1, but it seems to work reasonably well with other checkpoints. It does minor variations of a basic barefoot seated pose on the ground with the soles of the feet facing the viewer that every model I tested so far was unable to produce coherently.
Description
Completely overhauled training set to reduce biases towards certain faces and outfits:
No repeating backgrounds and faces (inpainted over every face and several backgrounds in the training set)
More diverse outfits (50% or more of the subjects were wearing denim jeans in the previous version)
More diverse poses (spread toes, feet touching/apart, feet up, knees flexed) but can be hard to control in generations
Includes clean and dirty feet (specify preference in prompt)
FAQ
Comments (9)
I've been looking for this basic pose but for 1.5. Do you think you'd ever try your hand at that? I'd love to have this for it!
Probably not for a while (likely never) since I don't use SD 1.5 anymore and I barely even know how to train LoRAs for SDXL.
I must admit I wasn't very impressed with SD 1.5 back in the day, but I also only ever used it through the official txt2img script, which had extremely limited functionality. I only really got hooked on Stable Diffusion with SDXL and Fooocus.
If you can give me a good sales pitch why SD 1.5 would be better for certain types of generations than SDXL (other than just being more lightweight for training and inference), I would certainly consider it though.
@F1_F2_F3_F4 I gave up on XL a couple months ago out of not being up to the highly refined standards of 1.5 for photographic human subjects at a level of fine detail and realistic face composition. The smoothness of skin, eyes too big and often mis-aligned, triangular skewing face shapes, weird bumpy/wavy lips, and lack of very effective controlnet or img-to-img conditioning guidance turned me off of it.
Granted, at the speed of development in SD, things are probably light years ahead now, I just haven't gotten back to it with any of the recent models. I have also been learning how to work with CFG and CLIP more in depth, along with some other factors, so maybe my methods were off. Probably time to give it another go, but these are my reasons to remain with 1.5 presently.
I honestly believe that XL cannot possibly exceed what I've achieved in pure photographic realism of human subjects in 1.5. Gens that can be indistinguishable from an actual photograph are as good as it gets, and 1.5 can do it. I like what XL does with complex prompting, creativity/artistry, and flexibility, but it just isn't what I have been looking for. I will definitely work with it eventually, I just haven't really felt the need as of yet. That's my perspective for what it's worth.
@MysteryWrecked I see, thanks for taking the time to elaborate. I'll keep it in the back of my head, but for now I'll be working on the training set before even considering other source models as I would guess there's a fair bit of extra potential there.
Holy... This looks FANTASTICLY good! Thanks!
it doesn't work , I don't see it on my Lora list. Do you think you can make other feet stuff ?
I only tested it in Fooocus and here on Civitai and it worked in both cases as it's just a standard SDXL LoRA, but it's not compatible with SD 1.x or 2.x in case that's what you're using.
And yes, I will almost certainly be releasing other feet-related stuff (hopefully not too long from now), but most likely only for SDXL.
Literally the only lora that works for feet on epicrealismXL
different foot textures would be awesome!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.