Finegrain Product Placement LoRA

Comparison with raw Flux Kontext

Inspiration

Our company is working on image editing, with a focus on lightweight specialist models able to deliver higher quality and more control that generalist models, while being cheaper.

What it does

Our LoRA makes it possible to insert products into photos, as if they had always belonged. To do so, it extends FLUX Kontext [dev] with multi-image input support.

How we built it

We wrote a full blog post detailing the training procedure: https://blog.finegrain.ai/posts/product-placement-flux-lora-experiment/

Challenges we ran into

Super hard to gather high quality before / after pairs.

Accomplishments that we're proud of

We delivered an open source proof-of-concept, with a public Hugging Face space.

What we learned

Consistency is critical in before/after pairs: e.g., objects must be in the exact same state in reference and ground-truth images (avoid lamp on vs. off, empty vs. full glass, ...)
Text-only prompts are insufficient for fine-grained spatial control (placement & scale) - also, automatic annotations often lack accuracy and/or clarity.
Leveraging FLUX Kontext’s visual cue capability with bounding boxes gives predictable placements.
Hard to avoid burning base model capacity, especially around textures.
Challenging to achieve strong relighting/perspective correction while preserving the subject with high fidelity.