sanjanachintalapati/easy-circle-flux-lora

Easy Circle FLUX LoRA

LoRA fine-tuned on FLUX Kontext model to add red dots to images.

Training Details

• Base Model: lrzjason/flux-kontext-nf4
• Task: Add red dots on specified objects in images
• Dataset: 7,000 synthetic image pairs (blank → blank with red dots)
• Training Steps: 20,100 steps over 5.7 epochs
• Training Time: ~20 hours on A100 GPU
• LoRA Rank: 16
• Learning Rate: 0.0001 with cosine scheduler
• Batch Size: 2
• Image Resolution: 512x512

Available Checkpoints

• checkpoint-5-20100/ - Final checkpoint (step 20,100)
• checkpoint-5-20000/ - Step 20,000
• checkpoint-5-19900/ - Step 19,900

Usage

from diffusers import FluxKontextPipeline
import torch

# Load base model
pipe = FluxKontextPipeline.from_pretrained("lrzjason/flux-kontext-nf4")

# Load LoRA weights
pipe.load_lora_weights("sanjanachintalapati/easy-circle-flux-lora", 
                       subfolder="checkpoint-5-20100")

# Generate
prompt = "Add red dots"
control_image = ... # Your input image
output = pipe(prompt=prompt, image=control_image)

Training Code

Full training code and configuration available at: https://github.com/sanjanachin/qwen-image-finetune