BachNgoH/flux2-klein-4b-overlap-ownership-lora

FLUX.2 Klein 4B — Overlap-Ownership LoRA

LoRA adapter fine-tuned on FLUX.2-klein-base-4B for count-preserving dense image generation with overlap-ownership attention zones.

Model Details

• Base model: black-forest-labs/FLUX.2-klein-base-4B
• LoRA rank: 16, alpha: 32
• Target modules: to_k, to_v, to_q, to_out.0, to_qkv_mlp_proj
• Training steps: 2000
• Training framework: PEFT + diffusers

Training Configuration

• Effective batch size: 4 (bs=1 x grad_accum=4)
• Learning rate: 5e-5
• Optimizer: AdamW
• Precision: bf16
• Features: GLIGEN-style layout attention + AIBL loss + Overlap-Ownership zones

Overlap-Ownership Zones (Idea 7)

Three-zone classification for overlapping bboxes:

• Winner zone: pixels owned by one bbox (strong positive bias)
• Contested zone: pixels claimed by multiple bboxes (moderate positive bias)
• Loser zone: pixels occluded by other bboxes (reduced bias)

Evaluation Results (1000 samples, val split)

Metric	Value	Baseline
Exact Match	14.72%	15.34%
±1 Tolerance	36.00%	—
±2 Tolerance	51.26%	—
MAE	4.285	4.22
RMSE	6.844	8.12

Usage

from diffusers import Flux2KleinPipeline
from peft import PeftModel

pipe = Flux2KleinPipeline.from_pretrained("black-forest-labs/FLUX.2-klein-base-4B")
pipe.transformer = PeftModel.from_pretrained(pipe.transformer, "BachNgoH/flux2-klein-4b-overlap-ownership-lora")

Framework Versions

• PEFT 0.18.1
• Diffusers
• PyTorch 2.x