Nekochu/myt5-large-SD-prompts

🎨 Stable Diffusion Text2Text

Architecture

myT5 1.2B

400k samples trained

🔍 encoder-decoder >2023 models

DistilBART

✓ Efficient

encoder-decoder

ModernBERT

✗ No seq2seq

encoder-only

ByT5

✗ Tokenizer

training failed

myT5

~ Limited 256 ctx

encoder-decoder

Run the Model

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
model = AutoModelForSeq2SeqLM.from_pretrained("Nekochu/myt5-large-SD-prompts", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("Nekochu/myt5-large-SD-prompts")

prompt = "### Instruction:\nCreate stable diffusion metadata based on the given english description. a futuristic city\n\n### Response:\n"
inputs = tokenizer(prompt, return_tensors="pt", max_length=256, truncation=True).to(model.device)
outputs = model.generate(**inputs, max_length=256, num_beams=5, early_stopping=True)
result = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(result)

🏙️

Cyberpunk City

SFW

Nikon Z9 200mm f_8 ISO 160, (giant rifle structure), flawless ornate architecture, cyberpunk, neon lights, busy street, realistic, ray tracing, hasselblad

🐉

Fantasy Dragon

SFW

masterpiece, best quality, cinematic lighting, 1girl, solo,

😈

Anime Succubus

NSFW

masterpiece, best quality, highly detailed background, intricate, 1girl, (full-face blush, aroused:1.3), long hair, medium breasts, nipples

Experimental Model

• Structured output tasks perform best.

SHUT UP!

🔬 Training Evolution & Alternative Attempts

torchtune Qwen3-0.6B →

decoder-only

pytorch-lightning Zamba2-1.2B →

hybrid arch

tokenizer failed google/byt5

byte-level