Llama-3.2-3B-ImageGen

Llama-3.2-3B — Image Prompt Generation (LoRA Merged)

This repository provides a LoRA-finetuned & merged version of meta-llama/Llama-3.2-3B, specialized for image prompt generation.
It is designed to create cinematic, detailed, and structured prompts for text-to-image models such as Stable Diffusion XL and Flux.

Note: This is a prompt-generation model, not an instruction/chat model. It is trained to produce concise, creative prompts suitable for diffusion-based image synthesis.

---

Model Details

• Maintainer: KavinduHansaka
• Base model: meta-llama/Llama-3.2-3B
• Model type: Decoder-only causal LM (3B parameters)
• Languages: English (prompt tags, stylistic descriptors)
• License: MIT
• Finetuned with: LoRA adapters, then merged
• Training dataset: prompt-gen-10k-flux-sdxl

Model Sources

• Merged model repo: https://huggingface.co/KavinduHansaka/Llama-3.2-3B-ImageGen
• LoRA adapter repo: https://huggingface.co/KavinduHansaka/Llama-3.2-3B-ImageGen-LoRA
• Training dataset: https://huggingface.co/datasets/KavinduHansaka/prompt-gen-10k-flux-sdxl

---

What’s Included

• config.json, generation_config.json
• Merged model weights (model.safetensors)
• Tokenizer files (tokenizer.json, tokenizer_config.json, special_tokens_map.json)

---

Uses

Direct Use

• Generate stylized, cinematic, or structured prompts for image synthesis models (Stable Diffusion, Flux, SDXL).

Downstream Use

• As a base for further LoRA finetuning on style-specific datasets.
• As a prompt generator inside T2I pipelines.

Out-of-Scope Use

• General-purpose chat.
• Safety-critical applications.

---

How to Get Started

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

REPO_ID = "KavinduHansaka/Llama-3.2-3B-ImageGen"

tok = AutoTokenizer.from_pretrained(REPO_ID)
model = AutoModelForCausalLM.from_pretrained(
    REPO_ID, device_map="auto", torch_dtype=torch.bfloat16
)

prompt = "Create a cinematic noir macro photo with film grain, 1:1 ratio, sharp focus."
inputs = tok(prompt, return_tensors="pt").to(model.device)

out = model.generate(**inputs, max_new_tokens=120, do_sample=True, temperature=0.5, top_p=0.9)
print(tok.decode(out[0], skip_special_tokens=True))

---

Training Details

• Training data: prompt-gen-8k-flux-sdxl
• Training method: LoRA with PEFT, adapters merged into base model.
• Precision: bfloat16/float16 during training.

---

Technical Specifications

• Architecture: LLaMA 3.2 (3B parameters)
• Hardware: NVIDIA GPU ≥6 GB VRAM
• Dependencies: transformers, peft, accelerate, torch, sentencepiece

---

Citation

@misc{llama3.2-3B,
title = {LLaMA 3.2 (3B)},
author = {Meta AI},
year = {2024},
url = {https://huggingface.co/meta-llama/Llama-3.2-3B}
}

@misc{llama3.2-3B-ImageGen,
  title  = {Llama-3.2-3B Image Prompt Generator (LoRA Merged)},
  author = {Kavindu Hansaka Jayasinghe},
  year   = {2025},
  url    = {https://huggingface.co/KavinduHansaka/Llama-3.2-3B-ImageGen}
}