Fares Fares - Flux.1 D

Fares Fares (born 29 April 1973) is a great Swedish-Lebanese actor of Assyrian descent. Known for his performances in the video game A Way Out (2018), TV series Westworld (2018), Chernobyl (2019) and The Wheel of Time (2021).

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 1.37

For close-ups and portraits, you can go with 1 or 1.1

Trigger Keyword:

Fares Fares

 Supporting prompts:

Age:25, age:35, age:40, age:45, age:50, medium-length hair, slicked back hair, grease hair, half-open mouth, closed mouth, stubble beard, beard, reading glasses, open-mouthed, side view, profile, front view

 Negative supporting prompts:

worst quality, Low rated, jpeg artifacts, blurry

Flux was not as great as SDXL with learning his appearance throughout the years/ages. Probably because I did not rewrite his captions. Flux would prefer "25 yeas old" I guess. Anyway, it's possible to make him look younger and older. But the standard 40-50 will be more prevalent.

This LoRa training come out great! The resemblance is unbelievable and fantastic compared to any of the SDXL and SD1.5 versions. It learned perfectly his nose size and shape, his ears shape, eyebrows, eye shape, wrinkles... well everything.... except for moles, birthmarks and tattoos.

I have no idea how to make Flux learn those details. He has a mole on his right eyebrows, skin birthmarks on his right cheek close to the eye, and a bunch of tattoos.

The tattoos were inconsistent on the dataset. On some images the tattoos were present, and other the tattoos were removed, so I didn't expect it to learn those.

Anyway, I had the hardest time choosing an epoch, it took me 3 weeks of testing... That is because from epoch 2 to 12 it was 10/10 with resemblance. Simply perfect resemblance. In the end, after endless xy plots, I choose epoch 7. I also choose a higher weight because it makes great full-body images keeping his face.

His body shape is well learned but not as good as SDXL learned it. I don't know why. I feel like SDXL really got his body proportions right. But this Flux is bot bad at all. It won't make the standard perfect muscular toned man, unless you ask for it.

I tried doing a block analysis and preset 5 was good. But in the end, epoch 7 is good enough and don't bring enough distortions to the model, so no need for adjusting blocks, IMO.

Of course, bleeding to the same class (man) is bad, as in any Flux LoRa (and even the base model itself). But it does bleed to women as well. I don't consider this a big problem. It's a side effect of pretty much any character LoRa. My block preset 5 does help a little, but I don't think it's worth it to publish it only because of this.

Text is well-preserved, but if getting too much garbled text, lowering the strength to 1.0 can help.

Problems with the current Lora:

• Moles, birthmarks and tattoos not learned.

• Bleeding quite bad into other people faces (this is expected, unfortunantly)

• Body shape is normally on point, but sometimes can be the standard muscular flux man. You can state his age to get better body proportion resemblance.

• Sometimes hair can get weird in a second pass upscaling. Lower the weight a little for those cases.

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

Just for my own reference later: epoch 7 with preset 5 (1.37); epoch 8 (1.2); epoch 10 (1.2); epoch 10 with preset 5; epoch 12 (1.2) and epoch 12 with preset 5 were all really great. Almost the same as the choose epoch.

Trained on Dev2Pro (don't generate on it, gen on normal dev). I used Prodigy. Trained don FULL FP16 on Kohya. Steps 5285 of 9060. Trained only on resolution 1024. No TE. No regularizations. Rank/Dim 16/16.

I included some more images (screenshots and CGI concepts) from his A Way Out. Total Images: 355. Some repeated cropped faces. High quality images were repeated 5x.