Hey ya! Long time no see uh?
Yeah, I wasn't dead just.. busy, really. A lot of stuff happened (in the right way) but well, still doing some testings over there and I think it could be cool to share this with the community. Let's try some fluxing, ^^.
I trained a Thanos FLUX-D LORA using different outfits, image styles, etc. since I found Flux itself had some problems (not always) to create the proper outfits (classic and cinema).
Still unsure about best caption system, training opt. but I think practicing enough I will figure it. Matter of time and trial and error I guess. So for now, that.
Not much to say, use your favourite prompting system and prompt:
Option A) Since model already recognize thanos (but not so good?), prompt "thanos" for better adherence at 0.4 onwards (if you want minimal changes from base 0.4 is ideal while improving the aesthetics, see the final image in the gallery uploaded).
Option B) For a new experience prompt "th4nmv" and make your tests.
Still I have some considerations to say.
If you have problems (I doubt it) use a longer prompting.
Skin failing? skin not purple enough? "purple" is your friend (purple skin, purple, etc.).
Want "classic" comic suit? "classic blue and yellow battle suit" is your friend.
Helmet also can be added with "helmet" or "classic blue and yellow helmet".
Infinity gauntlet is also a subtoken in the case you want to try.
As always thanks to everyone to support me and my friend in Patreon.
PS: Respect the license written here:
https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md
Thanks. This was trained for test and for fun. Not for commercial purpose or anything.
Description
1.0
FAQ
Comments (13)
Hey, this is amazing, I am also trying to train a LORA for dragon ball characters but am not able to get good results, may you share your training methods and training data, that would be a huge help
For now pretty basic stuff since we are right now in the darkness of captioning... trigger unique to test (even you can train as Thanos an prob better), around 50 images for this test, 2000-2500 steps and adamw8 optimizer (testing others btw).
@LDWorksDavid great, but I am confused for how to caption the images, I trained a character, 9000 steps, character is good but the problem is I am not able to alter the character but I can able to alter the Thanos, using your Lora, may you help me on captioning?
@krigeta captions I used were experimental. No rules followed, just describing what I was seeing like in SDXL for ex, but giving more room to details on the descriptions. You can even use GPT to autocaption the image and tweak them after. Also is probable that 9000 steps are a lot and you overfitted the LORA?
@LDWorksDavid yes I guess but it would be amazing if you only provide me the 4-8 samples you used for your training
@krigeta Samples? I don't have them anymore. The samples for validation were deleted after LORA finished to train and check.
@LDWorksDavid by sample I mean the 4 images + 4 txt files
@krigeta Ah, input data. Don't have anymore (they were MJ random Thanos images, some comic panels and some movie shots, mostly MJ btw) but I used GPT for captions, here is one:
The image depicts th4nmv. He has a bald head with a distinctive, ridged chin. The character is dressed in an intricate, gold-trimmed, blue and gold armor, which includes a chest plate and shoulder pads. This is his classic battle armor suit. His attire is complemented by large golden gauntlets on both hands, and he is shown clenching his fists in a pose that exudes strength and determination. The background and the setting suggests a sense of impending conflict or aftermath of a battle.
For example. As I said I mixed prompts and did experiments so can't help more than use GPT, tweak them and experiment cause right now I'm not so sure works good U^^.
Holy shit the star wars one goes hard
Working well for fp8_e4m
Flexibility: Mixes well with other loras and the strength is highly responsive
Concept: Easy to manipulate Thanos into different apparel, environments and ages.
Thanks for the review!
How about a comic / marvel style like in the first image ;)
Umm, I think maybe not worth too much since the model already can do comic style and if you reference it with some prompts to comic artist names you can pull it off but it's not a bad way of thinking... It's doable of course. For example, you can train some Jin Lee artworks with a mix of MJ or whatever and probably you can get something similar. And you can change Jim Lee for whatever that is flat line sharp edge drawing.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.