How to Train FLUX.1 Faces For <100credits
Used 14 images for my dataset, costed me 94.5 credits, results were😙👌NEWBIES GUIDE Step 1.DO NOT USE FLUX.1 as base.i used FLUX 2.8D Beauty - 1.0 https://tensor.art/models/769267843150826284i ran the LoRA on it's base and worked really well albeit looking a little differentbut when i ran it with FLUX.1 - dev-fp8 https://tensor.art/models/757279507095956705[NSFW] FLUX-成年人内容合集大模型Fp8 - fp8 https://tensor.art/models/780391846850510089 the results were shockingly realistic, quite literally a cloneand i only used 14 images because that's all i have.Step 2.prepare your datasetmake sure your images are aligned in your 512x512 or 1024x1024 imagesand i mean actually make a guide for where the eyes and mouth and nose should go in the imagesif you use photoshop, bring out the ruler, get a cross section line and place the face on it, with the cross on the eyebrow and nose(t-zone) and adjust from there. if not just get gimp or krita or whatever free software you can findmake sure the image is straight, like the end of the eyes are parallel to the guide and from there create line guides for the mouth and or any other parts you want to take note ofnow start dragging and shrinking/stretching the images to fit these lines, turn the image if you have to, if there are blank spaces at the edges/corners, just fill them out with white or whatever background color the image hadStep 3.start a new trainingthere isn't much to change from here, choose the base FLUX 2.8D Beauty, add your trigger words, now you want this part to be more than 1000 stepsso i just did 1 repeat and maxed out the epoch, which theoretically means you could be fine with 10 images(1x100x10=1000), but the more the better(also more expensive)personally i think more epochs is better than more repetitions, i could be wrong or maybe there's a different use case.saving every n epochs just means how many trained files you will get at the endi did every 10 just so i could see how the training progresses, but u can just take the last 10 if you want by making it 1 instead of 10 like mine (though some might say there could be "over trained" models which means undesirable effectsbut with such a small dataset, all my trainings showed the best results after 70 epochs and they stagnate there(basically 70 80 90 100 epochs all looked the same to me)Note that using this base will beautify your models quite literally, also it was the closest i got to likeness to the original when i paired with the actual FLUX.1 base.know that i tested training with all 3, and paired all 3 with each of the base(beauty>beauty | beauty>flux.1 | beauty>flux unlocked)(flux.1>beauty | flux.1>flux.1| flux.1>flux unlocked)(flux unlocked>beauty | flux unlocked>flux.1 | flux unlocked>flux unlocked)and the results from FLUX.1 and FLUX unlocked were mediocre, sure it works maybe with 100+ dataset but the output didn't manage to translate? the model's features that well, or maybe the base just wasn't a good fit for the face shape of the model.So in the end i just took my epochs 80 90 and 100 from FLUX 2.8D and saved as models with each of the base accordingly(beauty>beauty | beauty>flux.1 | beauty>flux unlocked)you can just save however many use cases as you needStep 4.train it. once done, browse around for some prompts u can test and try generate, see which results you liked the most.once again, this version beautifies your model, so you might wanna try saving one as a model and try generating with the base FLUX.1 to see if you like it. 👍TL;DR1. train using FLUX 2.8D Beauty - 1.0 https://tensor.art/models/7692678431508262842. prep, align, crop, rotate dataset images3. train to at least 1000 steps4. save one of it, test it with different base, see if you like the resultslmk if there are better ways? i was just shocked how easy and realistic it got because when i tried training with 100+ images before i got unusable results
