A DreamBooth LoRA for FLUX.2 [klein] 9B that learns a single specific product — the Air Jordan 1 Retro High OG "Chicago" (style 555088-101) — and renders it into new photographic scenes from a text prompt while keeping its silhouette, colourway and logo placement consistent.

It was trained as part of a product-imagery workflow: fit a small adapter on a handful of reference photos of one product, then generate unlimited on-brand studio and lifestyle shots of that product without a photoshoot.

Trigger

Refer to the shoe in your prompt with the full phrase:

tjkzx Air Jordan 1 Chicago sneaker

The phrase pairs a rare token (tjkzx) with the real product name. The token gives the adapter a clean, collision-free handle; the name lets the base model contribute its existing knowledge of the shoe. Everything else in the prompt describes the scene — background, lighting, lens, mood.

Usage

python

import torch
from diffusers import Flux2KleinPipeline

pipe = Flux2KleinPipeline.from_pretrained(
    "black-forest-labs/FLUX.2-klein-9B", torch_dtype=torch.bfloat16
).to("cuda")
pipe.load_lora_weights("manideep63/aj1-chicago-lora")

prompt = (
    "a product photo of tjkzx Air Jordan 1 Chicago sneaker on a white marble "
    "podium, soft studio softbox lighting, subtle shadow, e-commerce hero shot, "
    "85mm, sharp focus"
)
image = pipe(prompt, num_inference_steps=8, guidance_scale=1.0).images[0]
image.save("out.png")

FLUX.2 klein is a distilled model: keep num_inference_steps between 4 and 8 and guidance_scale around 1.0. Higher step counts add time without much benefit.

Repository contents

Path	Description
`pytorch_lora_weights.safetensors`	Final adapter, rank 96 (fp32)
`checkpoint-250 … checkpoint-2500`	Intermediate checkpoints, saved every 250 steps
`inference/`	Ten sample generations (default scene set)
`sophisticated/`	Twelve sample generations (detailed campaign-style prompts)
`logs/`	TensorBoard training logs

The final adapter at the repository root is the recommended weight. The numbered checkpoints are provided so you can trade a little identity strength for a little more scene flexibility if a particular prompt needs it.

Training


Base model	FLUX.2 [klein] 9B
Method	DreamBooth LoRA, rank 96 (alpha 96)
Reference set	16 images of the one shoe — 12 studio angles from a 360° set plus 4 lifestyle shots
Captions	Scene-only: each caption describes the surroundings and never the shoe's colours or materials, so identity is carried by the trigger phrase and the images
Optimiser	AdamW, learning rate 1e-4, constant schedule, 100 warmup steps
Precision	bf16 training, fp32 saved weights
Steps / batch / resolution	2500 / 4 / 1024
Guidance (training)	1.0

Intended use and limitations

This adapter is for product-photography style image generation of one specific sneaker. Identity is most faithful on the canonical white / black / red colourway in clean studio-to-lifestyle compositions. Very small, heavily occluded, or extreme-angle renders can lose fine stitching and panel detail. As a distilled model, klein follows the overall mood and composition of a prompt well but is less literal on very long prompts than the full FLUX.2 dev model.

License and trademarks

This adapter is built on FLUX.2 [klein] 9B and is distributed under the same FLUX.2 Non-Commercial License — non-commercial use only. See the linked license for full terms.

"Air Jordan", "Jordan", the Jumpman logo and "Nike" are trademarks of Nike, Inc. This is an independent research and educational demonstration of product-imagery fine-tuning. It is not affiliated with, authorised by, or endorsed by Nike, Inc. or Jordan Brand, and all rights in the product and its branding remain with their respective owners.