All Images below are sets of 3 different raw outputs (no inpainting, controlnet, etc.)
Flux_dev
t5xxl_fp8
Random Seed
896 x 1152
euler
20 steps
"Polaroid photo of a of a 35-year-old Korean man walking on a beach in the Caribbean. There are people swimming in the water. He is wearing a red t-shirt."
"Drawing of a 35-year-old Korean man walking on a beach in the Caribbean. There are people swimming in the water. He is wearing a red t-shirt."
"Amateur photo of a of a 35-year-old Korean man walking on a beach in the Caribbean. There are people swimming in the water. He is wearing a red t-shirt. The photo has some jpeg artifacts and there is slight overexposure."
I generated more images above because I think I found a "style" that has less consistent prompt adherence. 2/6 images above have very sharp focus and no jpeg artifacts as if they were professional photos. These were originally generated as a control group for the Amateur Photography LoRA
The next images were generated to make sure there wasn't some other kind of error causing this:
"Professional photo of a of a 35-year-old Korean man walking on a beach in the Caribbean. There are people swimming in the water. He is wearing a red t-shirt."
These images generated as expected, so let's start adding the LoRA to see if we get better style adherence:
"Amateur photo of a of a 35-year-old Korean man walking on a beach in the Caribbean. There are people swimming in the water. He is wearing a red t-shirt. The photo has some jpeg artifacts and there is slight overexposure."
LoRA Weight: 0.50
LoRA Weight: 0.75
LoRA Weight: 1.00
LoRA Weight: 1.25
Let's try with a new prompt:
"Amateur photo of a Facebook profile picture of a 25-year-old woman sitting in her living room, the room is cluttered. She is sitting in front of a large window and it is night time. The photo has some jpeg artifacts."
LoRA Weight: 0.00
LoRA Weight: 0.50
LoRA Weight: 0.75
LoRA Weight: 1.00
LoRA Weight: 1.25
Link to full resolution of all the images above.
Conclusion:
Everyone knows about FLUX prompt adherence when it comes to subject matter and text. The style adherence isn't bad but is not to the level of SD1.5 or SDXL, especially (obviously) for trained models and LoRA's on a specific style.
When adding some LoRA's to FLUX generations you do get a more consistent style adherence at the expense of some control over the image (note that the higher the LoRA weight the less "Korean" the subject looked).
This may or may not be more of a comment on the specific LoRA used rather than on FLUX. I suspect the community will continue to improve LoRAs for FLUX and it is a matter of time before FLUX will surpass SDXL on all fronts (except GPU efficiency).
Let me know your thoughts.