Type | |
Stats | 47 |
Reviews | (10) |
Published | Oct 27, 2023 |
Base Model | |
Training | Steps: 4,188 Epochs: 12 |
Usage Tips | Clip Skip: 1 Strength: 0.8 |
Trigger Words | Fares Fares |
Hash | AutoV2 89B068ED5A |
Use Lora weight 0.8
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).
He is a very handsome man with very distinctive face characteristics that made me think he would be a good first try to create a SDXL lora.
You will get the best results and resemblance with the prompt keyword (Fares Fares:1.3)
Keywords:
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:
Low quality, Low rated, jpeg artifacts, dim lit (as there were some pictures with bad quality but captioned as so on the dataset)
Note on keyword: I thought that because his name is a repetition, that only “Fares” would be enough as a keyword, but it is not! As the training data was caption “Fares Fares”, the two words together trigger the best effect.
I’m so obsessed with getting perfect results that I ended up doing 10 different trained lora versions and Textual Inversion to see if I could improve results using Textual Inversion together with the lora, but the improvement, if there is any, is minimal. So, I don’t really recommend but it can improve results sometimes to use my TI embedding at 0.6 strength.
Also, I could not decide on a perfect version, so I uploaded two. One is the recommended one more inflexible with a little bit more resemblance. The other version is a more flexible one (can easily change hair color, hairstyle etc and also keep a good resemblance).
SDXL do spit out great beautiful images and is way more naturally flexible (it can easily change his age within the captioned dataset but also beyond and below, like creating a young Fares Fares, for example). But SDXL was harder to finetune and get it on point with perfect resemblance. Just as a comparison, with his SD 1.5 I got great resemblance and results with my first training version.
Yes it’s a big sized lora, I know. But pruning it lost expressive detail quality. I trained some versions that was smaller, but they all lost in the comparison xy plot tests to the big ones. Maybe I’ll find better settings in the future.
Try out my SD 1.5 and TI here (I'll add the link as I upload them):
Fares Fares – TI SDXL (not sure if I'll upload this one. Not great)
Fares Fares – Textual Inversion SD 1.5
The base checkpoint is the sdXL_v10VAEFix 6.7GB. As of right now, I don’t recommend any particular finetuned model as they don’t give expressive better results comparing to the original one.
CFG: 5.5
Clip skip: 1
1024px
DPM++ 3M Exponential (50 steps or more)
DPM++ 2M Karras (25 Steps or more)
DPM++ 2S a Karras
I did not tried lora block weights yet, it might be what it is missing to make it perfect. Maybe in the future.
Want to have some fun? Install wildcards dynamic prompts extension: Here it is a good crafted script I made for xyz pot testing. Paste on prompt and have fun with crazy results:
<lora:Fares_Fares:0.8> {03::profile side view|front view} {0.3::extreme face close-up|0.3::face close-up|medium body shot|far away full body shot} (Fares Fares:1.3) age {50|45|40|35} {on the beach|taking a selfie on a mountain rock|0.5::laying down in bed|wearing a clown costume|wearing a cop uniform|wearing a soccer uniform|wearing open unbuttoned shirt}, {||0.3::pink|0.3::blue|0.3::blond} {short grease hair|medium-length hair|stylish quiff hair|shoulder length long hair|long tied back man bun hair}, {tank top shirt|t-shirt|shirtless|suit}, {long scruffy|(shaved:1.2)|stubble|goatee|mustache only} beard, {happy face|devilish face|laughing|serious face|pouting face|angry face}, natural lighting, {sunlight|0.5::moonlight|0.5::dark dim nigh|0.4::neon city lights}, 4k uhd, dslr, soft light, high quality, Fujifilm XT3
Current problems with the Lora:
Same problem as always with SD, bad hands, teeth, crazy body proportions and distortions. Don’t think it’s the lora problem itself.
Eyes look worse than normal SD unfortunately. I don’t know how to get better results. Probably need extreme close-up pictures of his eyes, which I could not find.
Some settings: Trained 1024 res. 177 pictures. 10 steps for great quality pictures, 5 steps for OK pictures, 3 steps for cropped faces. 3 different folders/concepts (sharing the same “Fares Fares” + person/face concept). Actually trained on prodigy this time and I think it is good for SDXL. I still prefer AdamW for SD15.
I should have used some of the CGI pictures from the game "A Way Out" to make it more flexible with nonrealistic results… maybe in a future version.
Hopefully you can leave some results and some comments. Any idea is appreciated. Thank you.