abhi10s/flux-1-dev-adforge-lora

AdForge LoRA: Synthetic Advertisement Generation

This is a LoRA (Low-Rank Adaptation) adapter fine-tuned on the FLUX.1 Dev architecture. It was trained to generate high-fidelity, commercially viable product photography and few advertisement visuals .

Trigger Word

To activate the fine-tuned style, you must include the specific trigger word in your prompt.

Trigger Word: TOKN_AD_STYLE

Example Prompt:

"A cinematic shot of a perfume bottle, TOKN_AD_STYLE, 4k, studio lighting."

Training Details

• Hardware: Trained on AWS EC2 (g5.xlarge) instance equipped with 1x NVIDIA A10G Tensor Core GPU (24GB VRAM).
• Framework: Fine-tuned using kohya_ss scripts.
• Base Model: black-forest-labs/FLUX.1-dev
• Optimization: Gradient Checkpointing enabled to optimize VRAM usage.

Hyperparameters

• Resolution: 1024x1024
• Precision: bf16 (Bfloat16)
• Training Method: LoRA

Usage (Diffusers)

from diffusers import DiffusionPipeline
import torch

pipeline = DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16)
pipeline.load_lora_weights("abhay10singh/flux-1-dev-adforge-lora")
pipeline.to("cuda")

# Notice the trigger word 'TOKN_AD_STYLE' is included in the prompt below
prompt = "A futuristic sneaker advertisement, TOKN_AD_STYLE, cinematic lighting"

image = pipeline(prompt, guidance_scale=3.5).images[0]
image.save("ad_output.png")