Rythmind / Loïc Barcourt (Beatboxer) - SDXL

Loïc Barcourt, 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
 

Trigger Keyword:

Loïc Barcourt

 Supporting prompts:

Person, young man, [blue eyes::0.15], best quality

 Negative supporting prompts:

Muscular, woman, low-res, worst quality, jpeg artifacts

The blue eyes won’t always come up, but it's consistent. Sometimes they might be deformed, but that is normally a SDXL problem, and I see that a block analysis normally worsen the eyes, but IMO it’s a good trade because it won’t interfere with the background or other details of the composition.

This another LoRA that took me forever. I am not fully satisfied with the results, but the resemblance is good enough.

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. Because of that, I did some ‘tricks’ like using more repeats on the quality folder. Still, it needed a lot of experimentation.

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.

I did NINE training sessions. I’ve tried low, medium and high rank. Unique token instead of his name (didn’t improve anything, I prefer to use his name). Different settings and a lot of epoch, weight and block testing.

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. There was no version of the trained LoRas that learned that, unfortunately.

It should not change your original character or composition, even at high weight. But you can lower the TE even further or even do a 0 weight TE if you want <lora:name:0.2:1>

The base checkpoint is the “sdXL_v10VAEFix” 6.7GB. So, it should be very flexible with any checkpoint.

For the smaller sized version (v8) I actually used juggernautXL_v9Rundiffusion.

As of right now, I recommend juggernautXL_v9Rundiffusion and juggerxlInpaint_juggerInpaintV8 for inpaiting. I’m also enjoying Mobius for general, flexible, artistic generations.

Lighting models works great! I recommend Dreamshaper SDXL

I prefer 6 steps with DPM++ 2S a Karras CFG 2.2 and high-res for 5 steps 0.42 denoise and 1.35x res. But the default is DPM++ SDE Karras, CFG 2, 4 steps.

The Juggernaut lighting is excellent, too. The hyper, not so much.

For standard generation:

Lower CFG works better

Clip skip: 1

1024px

DPM++ 3M Exponential (50 steps or more)

DPM++ 2M Karras (25 Steps or more)

DPM++ SDE Karras

DPM++ 2S a Karras

‎

Use Adetailer specially for far away compositions.

Problems with the current Lora:

• Block analysis sometimes makes worse eyes, but it’s not that different from the bad pupil eyes SDXL already do.

• Resemblance is not always great

• The mole below his left eye won’t be generated

• His eyebrow thickness are sometimes too thin

• His nose shape are sometimes not big enough

• His ears shape is always spot on, but sometimes are too small comparing to his head

• All dataset had a beard. It was all captioned, but it’s really hard for him to not have a beard.

• Hair color, beard length, is really not that flexible.

• Face expression was captioned, but it’s not that flexible

Block Weights. I know this is very confusing, but it’s for my future reference:

I did analyze the blocks weights. This is quite a chimera.

1-      Recommended version (165mb) is a merge of: (1.0) of v4 epoch 3 lbw=0.25,0.8,0.8,0,0,1,1,1.1,1.35,1.35,1.5,0 + (0.18) of v7 epoch 4 lbw=0.2,0.8,0.8,0,0,0.25,1,1.1,1,1,1,0 and then after the remerge I lowered back some blocks, just for my reference I’ll leave it here: v4Re+v7Re + 1,1,1,1,1,0.2,1,1,0.8,0.75,0.75,1

2-      The low rank smaller version is a little bit “simpler”: v8 epoch 10 lbw=0.6,0.8,0.8,0,0,0.2,1,1.1,1.35,1.35,1.5,0

Maybe these blocks can be applied to all other people/characters LoRas, I’ll sure try them on my Fares Fares loRa that I have not done a proper block analysis yet.

I did not rescale this LoRA. It works at weight 1.

Some more settings: v4 (main). Trained 1024 res. 230x3 total images. Three different captions, CogVL, Florence Full, Florence Finetuned, ALL with WD14 tags. Epoch 3 from 6. now prodigy 1.0. 1 steps folder repeat for bad images, 3 repeats for best. Constant BATCH 6, rank 24/1,Scale weight norms 5, snr gamma 5, Noise offset 0.0357, no regularization image, Max Token Length 225. Shuffle caption, keep 2 tokens the loric and the first natural description. dropout 0.1

Some more settings: v7. Trained 1024 res. 244x3 total images. Three different captions, CogVL, Florence Full, Florence Finetuned, ALL with WD14 tags. Epoch 4 from 6. now prodigy 1.0. 1 steps folder repeat for bad images, 3 repeats for best. 1 repeats for face close up on mole. Constant BATCH 2, rank 24/1,Scale weight norms 5, snr gamma 5, Noise offset 0.0357, no regularization image, Max Token Length 225. Shuffle caption, keep 2 tokens the loric and the first natural description. dropout 0.1

Hopefully you can leave some results and some comments. Any idea is appreciated. Thank you.