Have you ever wanted to see white or "occidental" women in impractically tiny Japanese cut out swimsuit lingerie? Or thicc women, women of color, or tatted women doing such a thing? While PredatorRat is doing the lords work dressing their flawless hyper-confident lingerie models in impractical non-clothing that's too small to move in, one must admit they have a type, so to speak. And this is where I come in.After much grueling quantitative research, wasted runpod instances, and stupidly accidentally overwriting a version that actually worked, I've taught AI how to make Japanese cut out swimsuit lingerie, I know, I know, thank you, gentlemen, please hold your applause, you're all too kind. With the power of AI, you too, can jam, I dunno, whoever the fuck?, how about "your mum"? into some Predator Rat t-back spats. Can I get a hooah?Trigger words are <code>spa t-back spats</code> you can throw in <code>blue</code> , <code>striped,</code> and <code>raglan shirt</code> if you want (the rationale was <code>spa-0032</code> is the product number for this particular item, I may retrain using this word with a more advanced captioning model such as "moondream2"). Maybe someday some one can tell me why prompting InstructBlip to include <code>t-back spats</code> into the captions made this work when the most precise and intricate descriptions of what it is actually in each picture failed.

Shimimaru×PredatorRat「T-back Spats with Line」〈BK×WT Line〉 紙魚丸×PredatorRat「Tバックスパッツ ライン入り」〈黒×白ライン〉

Used InstructBlip for captions with the prompt "Describe the woman's t-back spats, what else she might be wearing, what's in the background, and any other details that stand out in the image. Do not speculate about the image or mention what the image might imply or indicate. If there is something unclear about the image or you are uncertain about something, do not mention it."Used this adapted boilerplate code from the InstructBlip repo and the following settings:<pre><code>def generate_instructblip_caption(image_path):
	global prompt
	model, processor = model_manager.get_instructblip_model_and_processor()

	image = Image.open(image_path).convert("RGB") # Ensure the image is in RGB format
	inputs = processor(images = image, text = prompt, return_tensors = "pt").to(model_manager.device)

	with torch.no_grad(): # Disable gradient calculation for inference
		outputs = model.generate(
			**inputs,
			do_sample = True,
			num_beams = 8,
			max_length = 64,
			min_length = 1,
			top_p = 0.4,
			repetition_penalty = 1.5,
			length_penalty = 1.0,
			temperature = 0.3,
		)
	generated_text = end_trim_to_sentence(processor.batch_decode(outputs, skip_special_tokens = True)[0].strip())
	return generated_text</code></pre>Trained on content from Predator Rat's online catalog and some promotional material from Twitter.

Files with SHA256: 437689ee842309850ce172b5f7fdd7d6b1e5cd22ed40070ecd52201c8f8ebed0