Flux
Recommended strength is 1.
This version was trained (20, 24 images) on Flux.1 Dev. Picked the last model at 4000 steps. Trained on an A40 on runpod - took 3 hours.
I tried running it on schnell but it didn't work, if somebody can figure out how to run it on schnell/schnell-dev-merge, that'd be great - reach out.
SD1.5 v2
Recommended strength is 1.
This version was trained (20, 24 images) - (24, 16 repeats) on base SD1.5. Picked the 7th epoch. Trained with new workflow.
Pony
Recommended strength is 1.
This version was trained (20, 24 images) - (24, 16 repeats) on Pony Realism. Picked the 12th epoch.
The face likeness is pretty good as is, might not need the SD1.5 model at all with some trial and error. Use Euler A for the face detailer step.
SD1.5 v1
LORA trained on 70 publicly available images of Alexandra Daddario.
Trained on base SD 1.5, should work with most realistic models.
Use ADetailer if details are lost and strength of 1 should work on most prompts.
Description
FAQ
Comments (13)
All the celeb Flux loras I've seen so far have insane likeness, it's kinda mindblowing lol
Yeah man that is crazy. Don't know why we can't get that likeness on pony etc. Wish there will be flux for pony
@chrisss1 Stable Diffusion in general just cannot do a good likeness on the fly, even the likes of SDXL. Pony is just bad because it can't do specific people too well, the face and body are way off. Which is why with Flux it's super impressive because it's simply the lora doing its job, no Dreambooth from what I've seen.
@chrisss1 Pony doesn't have a LLM built in basically lol
@cosmoslayer26 that is true sadly..
How did you train a lora on flux, can you please share?
Trained on a A40 on runpod
GitHub - ostris/ai-toolkit: Various AI scripts. Mostly Stable Diffusion stuff.
There is now a guide for this (which is evolving): https://education.civitai.com/quickstart-guide-to-flux-1/#training-sd3
Not working for me with this models:
flux.1_dev_fp8_fp16t5-marduk191.safetensors
flux1-dev-bnb-nf4.safetensors
But thank you for your effort.
Kijai/flux-fp8 at main (huggingface.co)
it works with this
not sure about other variants
didn't work on schnell/schnell-dev merge either
@__2_ Ill try ^^ thank you :)
Looks amazing but why do all pictures come out a bit grainy? Almost looks like it needs to be upscaled?
The training data was 512x512 and there was a bug in the training tool at the time.
It was the first flux model I trained, was just testing the waters.
Will revisit later.
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.