Loïc Barcourt (born 19 September 1988), better known as Rythmind, is a French beatboxer and looper. He is a current member of Berywam. He does short videos for YouTube and TikTok with the phrase “Can you remake this with your mouth?”.
Please be responsible, this is based on the resemblance of a real person, so follow Civitai rules when posting. But please do test it! I’ll be glad to see some results.
Recommended weight is 1.0
V1.5 is out, read the "about this model" on the right panel.
There's no need for negative now.
You can check some comparisons here: https://civarchive.com/posts/8384839
V1 info (outdated):
This NEEDS negative with at least a "worst quality, jpeg artifacts, screencap,..."
That is because it learned pretty well all the bad quality on the amount of screencaps I used since I didn't adjust the increase in repeats for the high quality ones. (Ostris tool don't have the folder repeat thing Kohya have)
To use negatives on Forge set DynamicThresholding (CFG-Fix) Integrated to enabled, use CFG at 3.5 or more. On the settings Mimic Scale to 1. Threshold Percentile to 0.99. Mimic Mode half cosine up. Mimic Scale Min 0. Cfg Scale Min 0. Sched Val 0. Interpolate Phi 0.75. This is just my settings. Not the best ones.
Trigger Keyword:
Loïc BarcourtSupporting prompts:
HD high professional quality photo, best qualityNegative supporting prompts:
worst quality, jpeg artifacts, screencap, (blurry:1.2)Flux got the blue eyes pretty good. The pupils quality might sometimes suffer (all the image quality suffer).
I'm not super satisfied with this LoRA. It got close to resemblance, but it's not perfect and the learning of the bad quality kind of sucks. I'll probably have to do another version.
Same as SDXL
The dataset is problematic because only a few images are really great high resolution quality images. A lot are low-res and a lot are screencaps.
I also experiment with a different idea to try to make it more flexible. I triplicated the dataset, and for each I used a different natural descriptions captioner. CogVL, Florence Fine tuned and Florence base full. First tag was his name + class “Loïc Barcourt Person”. The second tag was the natural description. After that, I added WD14 tags. I pruned all of it a little.
His mole/skin detail below his left eyebrow is pretty much impossible to generate without inpaint or other interferences. I tried creating a folder specific with that focus, but the zoom degraded the quality too much, so I could not use too many repetitions.
The base checkpoint is the “Flux dev original fp16 one”. So, it should be very flexible with any checkpoint.
For standard generation:
USE negatives, so CFG > 1
Euler + simple
50-60 steps is way better for photos on Flux
Problems with the current Lora:
Will decrease overall image quality to something like a screenshot. This sucks. Use negatives.
Resemblance is not always great
The mole below his left eye won’t be generated
Need more testing to see what else....
I wish I could do Block Weights on Flux...
Hopefully you can leave some results and some comments. Any idea is appreciated. Thank you.
Description
I admit this looses some resemblance, but the quality jump is unbelievable!
This is not a new trained version. It's the same as v1 but converted to kohya format and then I made a post training block analysis and remerged the block weights with my own tool here: https://civitai.com/models/880106/flux-block-weight-remerger-tool
With this preset here: 0,0,1,1,1,0,0,0,0,0,0,0,0,0,0,1,1,0,0,1,1,1,1,1,1,1,1,1.15,1.15,1.15,1,1,0,0,0.25,0.25,0.25,0.25,0,0,0,0,0,0,0,0,0.5,0.25,0.25,1.5,1.5,1.25,1.25,0,0,0,0
You can check some comparisons here (this is the "9" preset): https://civitai.com/posts/8384839
Loïc still deserves a v2 with a better dataset and training parameters, though.