Zoot's Human Photo Realmaxxer For Flux
FLUX1D lora
·
Uploaded Sep 25, 2024
·
Used 1.5K times
Trigger words
-
About this version
Expanded dataset.
About this model
A Lora designed to maximize photographic realism with regards to images of people, without sacrificing any quality and without needing to lower guidance below the stock setting of 3.5 (which has downsides such as loss of overall detail and reduced color range).
Trained onsite at 1024px / Dim 20 on an extremely diverse dataset of images that were all hybrid-captioned with both JoyCaption and wd-eva02-large-tagger-v3.
Recommended strength is between 0.7 and 0.9 or so, but feel free to experiment as you see fit.
All primary sample images were generated at stock 3.5 guidance with just Flux Dev FP8 and this Lora.
If remixing any of them, note that the strength used was NOT actually 1.0 in any case, CivitAI just doesn't handle ComfyUI metadata properly unfortunately. Set guidance back to 3.5 and try 0.7 strength as a starting point to get similar results.
Lastly, note that the dataset for this Lora did not focus on nudity, I intend to release a separate Lora for that which will be able to run alongside this one in a complementary way if desired.
This one isn't. V1 is trained on 276 hybrid-captioned (meaning NLP + tags) high res photographs of people of all ages and ethnicities, at native 1024px for 50 epochs with Prodigy and 20 Dim / 20 Alpha (in Kohya scaling). Future versions will likely expand the dataset further, but I'm pretty happy with this one as it is.
Recommended strength is around 0.7 to start, but feel free to experiment. This Lora however DOES NOT *need* to be run at full blast 1.0 strength, NOR SHOULD YOU start at 1.0 for any particular image as it is often in fact too high in a noticeable way. There is no particular "trigger word", just prompt as you normally would.
This Lora is trained at standard "SDXL equivalent" resolution buckets: 1024x1024, 832x1216, and 1216x832 are the most optimal resolutions for it.
Related Posts