** - First attempt / Please don't expect perfection. Should work in both Comfy and Forge. Mainly posting so people have a likeness model to test with their workflows/settings. There's a lot of variability at the moment (guidance? Sampler? fp8/fp16? etc) - so something to use to dial things in more perhaps.....
Trigger is "arnold person"
Trained on 12images of the Arnold for 1500 steps using Ostris's script (https://github.com/ostris/ai-toolkit?tab=readme-ov-file#flux1-training).
Use with the Dev model but may work somewhat w/ schnell and/or any merges.
Description
FAQ
Comments (9)
The model looks really cool! I'll test it as soon as possible!
What GPU do you have?
I tested the script with my RTX 3090 all night, without success. I know there are still improvements to be made to the script. But it's just to know if the problem comes from me or if I have to wait a little longer for an upcoming update?
the lora works very nicely! great job!
how long does it take to train compared to SDXL?
I was curious to undestand what Flux can do, and I'm happy to see the resemblance is strong! SD XL, on the other hand, has never convinced me about celebrities. It was crystal clear in graphics but always imprecise in proportions and faces, it barely did better than the SD 1.5
why it takes ages to making a single picturs ?? rtx 3090.. printscreen
I wish I could make FLUX loras :(
I've been trying to get the results you're getting, using the workflows from the image meta, and nothing is producing images with Arnold in them, its just random faces.
How is it possible to be nostalgic about movies that don't exist?
Arnold with his watch in any occasion