Version 2.0 update:
after I connected the grid pattern sometimes visible in bright or dark parts of the images with the fp4 training with OneTrainer, I switched to fp8 training, which got rid of the pattern. But I have to redo all my LoRAs to fix them all. Nothing changed, except the tag is now <lora:LeaningBack-step01200:1>.
After I discovered that Flux won't do certain poses with prompting, even if they are absolutely SFW, I tried to fix that. In this case if the model is leaning back, supporting herself with her arm. The trigger word is "sitting" and then "leaning back, supporting herself with her arm" (or arms). And for more consistency, a <lora:LeaningBack-50:1> tag.
In the set of version 2.0 are front angle pictures, but the model tends to side angle as these are the most common for this pose. Same with leg positions: bent sit is the most common, half bent sit and bent figure four (one leg bent and the other crossed under it) also exist, but are not reachable by prompt. It was difficult to decide which epoch was best - more give more control, but also more extra limbs. As always, a compromise with 50. I am still learning, and the variations for number of steps and images needed are big, but it seems like the 10 to 15 images for some of the leg positions aren't enough (the set has a total of 57). The question for the future is: either split the LoRA further and train the leg variants separately, or make extra LoRAs for different leg positions (and perhaps camera angles etc.), which will result in a kind of modular system.
Something I started here but had to be placed too in separate LoRAs to work is a finer control over face angles, body angles, camera angles and image area. But starting with this one, the captions are consistent as a preparation, even if they do not (yet) work standalone. The scheme is:
Face angles:
Full face view | Three-quarter face view | Two-third face view | Profile face view
Body angles:
Front angle shot | Three-quarter angle shot | Side angle shot | Three-quarter rear angle shot
Camera angle:
High camera angle | Eye level shot | Low camera angle | Shoulder level shot | Cowboy shot | Ground level shot
Image area:
Full-body portrait | Three-quarter-body portrait | Half-body portrait | Quarter-body portrait
It would be great if I get this to work!
(for version 1.0)
I had to use uncropped images for training, so there is a bit of impact on faces and other parts of the image. It should work with prompts like "she is leaning back, supporting herself with her right arm" or "she is laying on her left side, propping herself up on her left elbow".
Strength seems to work from 0.8 to 1.0 - if you get deformed hands or feet it is too strong.
描述:
This version has a somewhat different image set and captioning. The "leaning to the side" images are removed and moved to a separate LoRA. Then are the captions for face and body angles etc. added.
训练词语: leaning back, supporting herself with her arm
名称: LeaningBack-50.safetensors
大小 (KB): 25641
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success
