Product-Photo Accuracy Playbook

The root cause (validated)

Every full-generation model redraws the entire product on every call — drift is inherent, not user error. A 2026 arXiv benchmark documents that instruction-based editors (Google Nano Banana, Qwen-Image-Edit, HiDream-E1) systematically fail product identity: altered shapes, distorted text, changed colors, inconsistent branding. gpt-image-2 and FLUX Kontext live in the same bucket. Even Google DeepMind concedes that background swapping (compositing) is "the most detail-preserving method" — they pursue fine-tuned generation only because compositing can't relight or occlude. Conclusion: reserve redraw models for shots that truly need them; default to pipelines where the product pixels never pass through a generator.

The three paradigms

Compositing

Default — staged/hero shots

Segment the real photo → generate only the scene → add shadow → composite. Product pixels untouched.

Bria on fal (our account): fal-ai/bria/product-shot — $0.04/img, text or reference-image scenes, placement controls.

Bria native API: 6 placement modes incl. exact coordinates + 16:9 · dedicated Product Shadow endpoint · product/integrate embeds the cutout INTO a scene (the "tape on the wall" case).

Photoroom API v2: whole chain in one vendor (~5s/img): remove bg → AI Backgrounds (+seed) → AI Shadows → resize.

Traps: Bria defaults to ~1MP downscale — set placement_type=original + original_quality=true or dimples/layers die · integrate coords must match cutout aspect ratio (else stretch) · Bria result URLs expire in 1 hour · Photoroom: don't stack AI Shadows on AI Backgrounds; its "Relight" is exposure fix, not scene-matched.

Identity-locked generation

In-use shots only (unroll, peel, apply)

When the product must be redrawn, lock identity first.

Per-product FLUX LoRA on fal: flux-lora-fast-training — $2/1000-step run, is_style=false (subject segmentation); or flux-2-trainer $6.40/1000 steps, 9–50 photos, deploys straight to FLUX.2 inference. Train once per product, reuse forever.

Finegrain insertion LoRA (open weights, FLUX Kontext) — verified lessons: bounding-box visual cues beat text prompts (~15% of text placements failed) and rank-8/16 LoRA beats full fine-tuning on subject preservation.

Trap: LoRA fidelity on fine textures (dimples, layer stacks) is unmeasured publicly — bench before trusting. Complex subjects likely need more than the 9-image minimum.

3D-assisted

Multi-angle sets + video

Real photos → 3D asset → Blender scene → render exact geometry from any angle → optional low-denoise polish.

fal-ai/hunyuan3d-v3/image-to-3d: $0.375–$0.90 per product, takes front + optional back/left/right views (we already shoot these), outputs textured GLB → Blender.

Why it matters: one asset kills drift on every future angle and animates in the install videos. Directly answers the "might as well model it manually" ROI question — for $0.38 and minutes, not days.

Trap: mesh quality on layered edges / fine dimples is unverified — this prices the attempt, not the result. Verify against reference before adopting.

Decision table — per shot type

Shot type	Primary pipeline	Fallback	~Cost
Macro / material detail	Use the real photos. Don't generate.	—	$0
Hero on a surface (roll on deck)	Bria product-shot (fal) w/ `original_quality=true`	Photoroom chain	$0.04
Product INTO a scene (applied strip on wall)	Bria `product/integrate` (native API)	Per-product LoRA gen	~$0.04–0.10
In-use action (peeling, unrolling, hands)	Per-product FLUX LoRA + bbox-cue insertion	gpt-image-2 (current method) + human QA	$2 once + pennies/img
Multi-angle set / repeat product	Hunyuan3D → Blender render	LoRA	$0.38–0.90 once
Anything for the install video	Hunyuan3D asset animated in Blender	—	same asset

What did NOT survive verification

"Composite + relight" ComfyUI templates are fake compositing — the official Seedream 4.5 template feeds product pixels straight through a generative model (verified by downloading the workflow JSON: no mask/paste-back node). Same drift bucket. Avoid for accuracy-critical work.
Relighting chains (IC-Light v2, SwitchLight), practitioner case studies, in-use inpainting recipes, and QA automation (masked SSIM/DINO, VLM-judge): nothing survived — build and bench in-house.
No published benchmark tests reflective stainless / industrial products. Vendor quality words ("high integrity," "ensures lighting") are marketing until we bench them.
Refuted (0–3): Photoroom's cross-vendor pricing table — recheck all prices at purchase time.

Bench plan — ~$10, one afternoon

1 · Bria product-shot (fal) on the Butyl roll: placement_type=original, original_quality=true, deck/rooftop scene prompts. ~$0.50 for a dozen.
2 · Bria product/integrate (native API key needed): embed the applied-strip cutout into a real joint photo at exact coords. ~$1.
3 · Photoroom trial: same shots through remove-bg → AI Backgrounds → AI Shadows. Free trial credits / ~$2.
4 · Hunyuan3D v3 multi-view on the tape (we already have the P027xxxx angle photos) → import GLB into Blender, render 3 angles vs reference. ~$0.90.
5 · FLUX LoRA $2 fast-train on the product photo set → generate the peel/unroll shots that keep failing.
Score all five against the real photos (side-by-side page like MODEL_TEST.html) and adopt winners per shot type. Add a VLM-judge checklist (texture? width? layers? orientation?) so future QA isn't Benito's eyeballs.

Key sources

fal-ai/bria/product-shot · Bria product endpoints · Bria product/integrate · Photoroom API · Finegrain placement-LoRA experiment · fal FLUX LoRA trainer · fal FLUX.2 trainer · fal Hunyuan3D v3 · ComfyUI Seedream template (negative finding) · arXiv: instruction editors fail product identity

Bezar Video Studio · internal reference · verified July 2026 · pairs with the Polyguard Visualization Bible (art direction) — this doc covers the how, the Bible covers the what.