FLUX Klein Jewelry Photography Pipeline¶

Production pipeline for generating and compositing jewelry product photography using FLUX.2 Klein 9B. Covers cutout extraction, lighting manipulation, scene generation, and compositing.

Full 7-Step Pipeline¶

1. Original jewelry photo (client)
2. BiRefNet cutout → transparent PNG
3. IC-Light V2 Delight LoRA → remove original lighting
4. Klein 9B scene generation → background scene
5. Manual composite in Photoshop/Figma → rough placement
6. Klein 9B inpaint + Consistency LoRA → seamless blend
7. IC-Light V2 → final relighting to match scene

Step 1-2: BiRefNet Cutout¶

BiRefNet provides state-of-art transparent background extraction for jewelry. Handles thin chains, prongs, and edge detail better than older U^2Net.

from transformers import AutoModelForImageSegmentation
import torch

model = AutoModelForImageSegmentation.from_pretrained("ZhengPeng7/BiRefNet", trust_remote_code=True)
model.eval()

# Returns RGBA with alpha = foreground mask

Jewelry-specific challenges: - Thin chains: use highest resolution model variant - Diamond prongs: edges are sub-pixel thin at normal viewing distance - White gold on white background: pre-enhance contrast before segmentation

Step 3: IC-Light V2 Delight¶

Remove existing lighting to get a diffuse/ambient-only version before recompositing.

IC-Light V2 (ICLR 2025): - Flux-based relighting model (not SD 1.5 original) - Uses 16-channel VAE for light-aware latents - Non-commercial license (research and personal use only) - VRAM: ~18 GB for full-quality inference

from diffusers import FluxControlPipeline
from ic_light_v2 import ICLightV2Processor

# Delight mode: use fg-only model
pipe = FluxControlPipeline.from_pretrained("lllyasviel/ic-light-v2-fg")
delighted = pipe(
    prompt="product photo, neutral lighting",
    image=jewelry_cutout,
    num_inference_steps=20
).images[0]

API alternative (production):

import fal_client
result = fal_client.run(
    "fal-ai/iclight-v2",
    arguments={
        "image_url": jewelry_cutout_url,
        "prompt": "product jewelry photo, uniform soft studio lighting",
        "lighting_preference": "neutral"
    }
)

Step 4: Klein 9B Scene Generation¶

Generate background scene matching jewelry style and brand mood.

Prompt structure for jewelry scenes:

[Surface material] with [texture detail], [lighting type] from [direction],
[depth of field], [atmosphere], shot on [camera/film type]

Example prompts by jewelry type:

# Fine jewelry:
"dark velvet surface with subtle fabric texture, soft diffused studio lighting
 from above left, shallow depth of field with bokeh background, luxury editorial
 atmosphere, shot on medium format"

# Fashion/statement:
"rustic marble surface with grey veining, dramatic single-source rim lighting,
 deep black background gradient, high-contrast editorial, 35mm film grain"

# Bridal:
"white silk fabric draped softly, natural window light from right, cream tones,
 ethereal out-of-focus floral background, romantic editorial"

Step 5-6: Compositing + Inpaint Blend¶

After rough Photoshop composite:

Klein 9B inpainting settings for seamless blend:

Model: FLUX.2 Klein 9B (base or distilled)
LoRA: Consistency LoRA (consistency-edit-lora)
Denoise: 0.5-0.7 (higher = more creative blend)
Mask: expanded 20-30px around jewelry edges
Prompt: describe the shadow and reflection that should appear

Why Consistency LoRA: maintains the jewelry's original appearance while harmonizing its integration with the generated scene. Without it, Klein sometimes "redesigns" the jewelry during the inpainting pass.

VAE color shift warning: FLUX Klein's VAE decoder sometimes shifts jewelry metal colors (gold → slightly greenish, silver → slightly blue). Correct in Photoshop after compositing: use Curves on affected channel, +5-10 on warm tones for gold.

Step 7: IC-Light V2 Relighting¶

Remap lighting to match the scene's light source.

# Use fg+bg model for relighting with scene background
pipe = FluxControlPipeline.from_pretrained("lllyasviel/ic-light-v2-fg-bg")
result = pipe(
    prompt="jewelry product photo with [scene lighting description]",
    fg_image=composited_jewelry,   # jewelry on background
    bg_image=background_scene,     # reference for light direction
    num_inference_steps=20
).images[0]

Lighting prompts that work well: - "Soft window light from left, warm afternoon sun, slight shadows" - "Studio three-point lighting, soft fill, cool-toned background light" - "Candlelight from below, warm orange rim light, dark romantic"

Jewelry-Specific Challenges¶

Official FLUX.2 Klein LoRAs for Jewelry¶

These are BFL official LoRAs, not community-made. Load via standard ComfyUI LoRA loader.

Dataset Considerations¶

For training a domain-specific jewelry generation LoRA:

Caption strategy for jewelry: trigger word + scene description + material description. NEVER include lighting or style descriptors (those go into LoRA weights).

Comparison with SaaS Tools¶

Gotchas¶

• IC-Light V2 is non-commercial: the model weights are research-use only. For commercial product photography pipelines, use via fal.ai API which handles licensing.
• BiRefNet requires minimum 512px input: at lower resolutions, edge quality degrades significantly for thin jewelry elements. Upscale input if needed before running BiRefNet.
• Klein 9B edit model vs generation model: for Steps 4 and 6, use the BASE (non-distilled) Klein 9B. Distilled performs poorly for scene generation at the quality level needed for product photography.
• Consistency LoRA scale: the Consistency Edit LoRA is calibrated for scale 0.7-0.9. At 1.0, it over-constrains and prevents the model from adapting the jewelry to the scene.
• Manual composite is not optional: automated placement in Step 5 produces unnatural positioning 40-60% of the time. Human judgment for placement angle and position is still required.

See Also¶

• FLUX Klein 9B Inference - Klein settings and LoRA stacking
• skin retouch pipeline - frequency separation for texture preservation
• object removal inpainting - background cleanup
• Tiled Inference - high-res output for product photography
• synthetic dataset pipeline - building training datasets for domain LoRAs