So why am I uploading this as a new lora not as a V2 to my previous one ? simple I have done a thing....Don't ask me how I did It I have absolutely no clue! more below anyway :
If this version 4650 steps is too high I will upload a lower step version it will depend on feedback, And all sample images are first or second renders no cherry picking.
V2 of my blackwear slightly increased dataset, removed a few images, re-cropped. It likes higher steps for the text, however the text will not be perfect every time sadly. BUT What thing did I do ? in case you haven't worked it out yet from the pictures I accidentally allowed you to specify the text on the underwear it seems to work decently using flux dev, t5xxl_fp16,clip_I and ae.safetensors on forge, I tested eular simple/normal/beta all seem to work. I have tried strength 0.9-1 I am not sure if outside of this will work feel free to play ! it also seems to have maintained a good compatibility with character lora's. HOWEVER sometimes It needs things in the prompt to remove any bleed across such as "isn't wearing jewlery,she is only wearing a sports bra and thong" Im not sure how often you need this.
To use it use the phrase "Wearing a [colour] sports bra and thong, with text reading "text." on it" I'm not sure if separating this from your prompt with a comma or full stop will help it I haven't tested it extensively. to that end feel free to leave feedback how it works for you, if you find any better settings than what I post or notice anything I missed :)
Ok ramblings aside what did I change? well I changed the tagging to "wearing a sports bra and thong, with the text "blacked." on it", my idea being if I train a token such as "bl4cked" flux despite being smart might go "oh I dont know "bl4cked" maybe I should train body" so as a test I instead thought oh why not train it on something it already knows but I want to alter and well it seemed to work fairly well! the text being changeable seems to be a happy accident but if anyone is training loras you might find this interesting.
Description
FAQ
Comments (11)
A defining achievement from a refined gentleman... bunny
My thanks good sir
We are proud of you, perv
high praise
Slow clap
The power of flux... I'm find this spits out the correct text with even higher percentage than the v1 - and on top of that you can specify the text (and colour). I found that at strength 1 it was changing composition/character qualities a little too much but 0.9 has been working well. Still playing around with steps/schedulers etc. mostly Euler and 25+ steps. Had text come out perfectly at 20 steps at the beginning so who knows. Great work!
Glad its working ! yeah Its crazy how just a small change to the captions and a very minor tweak to the dataset has changed the results so damn much, flux op !
Nice work! I've been training similar models, for example for Nike Pro leggings and shorts, and I have the same sort of thing with the text going on. However, I will say that after training a few models and testing, that actually, the model was better when I didn't include 'text: "_____" on it' in the captions. What worked best was actually just training with ONLY a simple unique token. For example, I would just do "photo of a woman wearing black n1k3pr0 leggings, wearing sports bra, full body shot, plain background" that kind of thing. when I specified as well the text that was to be written it kind of made it less flexible somehow. To be honest, when I just trained it without that, and actually specified the text I wanted only when I was actually at the point of generating images, in my prompts, it would respond way better, and I could even change the text to whatever I want too! Feel free to check that out if you want to have a test of my model. I think we could learn some things from each other maybe :)
v1 was trained like that and its flexibility isn't the best, it worked with with character loras but as I think I put in the description I was getting body bleed, the way I captioned this one allows for flexibility, no body bleed, and any text. Its not 100% but its at a level where most people will be happy.
Could I ask you what parameters work best for you for this kind of concept, clothing with text? Especially the Network Dim and Alpha would be very useful for me, and the learning rate. Thank you :)
Feel free to join pyro's discord we discuss a lot of the stuff there.
@Gentlemanbunny could you get me a link? I've tried joining before but the links were all dead :/
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.