Comrade's Flashing Tits v2
I’ve doubled both the dataset size and the training steps; this is a massive leap forward from Version 1!
To fix the quirks of V1 and boost flexibility, I expanded the dataset significantly. But here is the hard part: I used an LLM to generate detailed captions for every single photo and then manually verified them one by one. (Honestly, I get bored just thinking about that process again). The result, however, is a Lora that performs leagues better than the first version.
General Impressions & Prompt Guide:
No Trigger Word: You don't need a specific trigger word. You just need to express the action you want using natural language. LLMs are great at helping with this, and you can check my sample images for inspiration. Actions like "lifting," "lifting up," "holding up," or "pulling down" now work with a high success rate. I made sure not to 'cherry-pick' the sample images too much, most are just results from my first few attempts.
Size Control: Adjectives describing breast size like Small, Medium-sized, and Large work much more accurately now. However, please remember that size is 'relative'... You might need to reroll a few times to get exactly what you envision. Based on my tests, descriptors like Flat chested or Huge also work, though perhaps with a slightly lower hit rate than the standard sizes. Subjective terms like natural, perky, or round can also help get the specific look you want.
Details: Some adjectives for nipples and areolas are also working. While not as consistent as the size tags, color descriptors like Light pink, Pink, Light brown, or Dark brown work well. Shape descriptors like Small, Large, or Prominent also work in many cases.
Coherence: I haven't observed any drop in the model's prompt coherence. I intentionally generated the sample images using very detailed captions to demonstrate this stability.
Any feedback you can provide after testing the LoRa will be incredibly helpful for planning my next steps. Hope you have fun! ^^
Version 1:
As a hobbyist who strictly trains character loras for personal use, I have to say I was really impressed by how 'trainable' Klein is. Seeing that potential gave me the courage to create my first lora to share with the community.
As you know, Klein is amazing in many ways, but it can be incredibly stubborn about generating anything 'spicy.' So, in my spare time, I put together a dataset and tried to train an AI image generation model to do what is honestly a 'must-have' capability. cough cough
Although there isn't a strict trigger word, I used the phrase "Flashing Tits" in all the captions, so you can include this phrase in your prompts when generating images.
It works well at a weight of 1.0. It doesn't drastically change the background or the subject; I'd say it's pretty stable. However, for most compositions, you can lower the weight to 0.7 or even 0.5. I managed to use it with my own character LoRa at 0.5 weight without altering the character's physical traits. Also, keep in mind that it works for in-painting edits as well.
I trained this using Ai-Toolkit with 520 hand-picked photos. To ensure high variety and adaptability, I made sure to select photos featuring different breast sizes, outfits, and actions.
Based on my limited testing, I can confidently say that—for now—this is the best solution for this specific action. But it is still far from perfect. Please keep in mind that this is my first lora and I don't claim to have any 'expert' knowledge. Here are some issues I've noticed:
Shirt mechanics: Pulling the shirt down, lifting it up, pulling it aside... I don't think I have full control over this yet. It kind of does whatever it feels like. I'm not sure if this is due to my captioning method, but I hope to improve this in V2.
Size consistency: Big, medium, small... sometimes I don't understand how it decides. Size is still a bit 'relative' for the model. I will try to refine this in the second version as well.
Bad news: No piercings.
I am grateful in advance for any feedback you can provide regarding the LoRa. Enjoy!