Product / Object LoRA from One Reference Photo (Angles, Materials, Captions)

Most LoRA advice assumes a face. Ecommerce sellers, prop artists, and brand marketers usually start with one catalog photo of a shoe, bottle, or weapon and need the product to stay consistent across scenes. This guide is the object sibling of creating a dataset from one image: when an object LoRA is the right tool, how to plan an angle / material / lighting grid, how captions differ from character work, and how to curate for shape consistency before you train.

What this guide covers (and what it does not)

Scope is building an object-consistent, captioned training set from one product or prop reference. Out of scope: Kohya folder trees, deep overfitting diagnosis, anime WD14 taxonomy, and a full base-model picker. You will leave with a decision table, a copyable product diversity grid, caption before/after examples, and a curation checklist.

Packaging after export: Kohya / Civitai packing. Baked white backgrounds or locked hero angles: overfitting vs undertraining. Base choice: Flux vs SDXL.

Object LoRA vs character LoRA

This post deepens the object/product row on how many images a LoRA needs.

Image count for products (cite, do not invent)

• This site: object or product 15–30—rigid subjects need fewer samples; cover every side and two lighting setups.
• Community starting ranges (labeled as such): Everypixel Workroom cites 15–30 photos of the same product with angles, lighting, and clean/simple backgrounds; Offline Creator cites simple object/product 15–35.

Coverage beats count. Ten near-duplicate hero shots are not the same as 15–30 useful sides. Plan to the 15–30 site range and fill empty grid cells before adding duplicates.

One strong product reference

• Resolution: prefer ≥1024 px on the short side (same native class as Flux / SDXL training).
• Silhouette: product not tiny in frame; defining parts visible.
• Lighting: even light that reveals material (gloss, weave, wood grain, metal edge)—not crushed shadows that hide geometry.
• Angle: three-quarter or clear front that shows depth; a pure top-down crop alone is a weak sole reference.
• Avoid: heavy compositing, watermarks, mixed SKUs in one frame, occlusion of defining parts, extreme filters.

Lock an object spec before you expand: silhouette, proportions, materials, colors that are part of the SKU, and unique markings the LoRA should own.

Diversity grid for rigid objects

Axes differ from the character grid: no expression or wardrobe; add materials, scale, and state instead.

Example cells across three SKUs

Filling every cell by hand is tedious. LoRA Dataset can expand one reference into object-consistent variations across angle, lighting, and background (free with your Gemini API key). Use it after you understand the coverage you need.

White background vs lifestyle (use both on purpose)

• Clean / white / neutral: teach silhouette, edges, labels, and proportions with less clutter.
• Lifestyle / in-context: teach scale, placement, and scene integration so gens do not only look like packshots.
• Only white-bg failure: hard to place in busy scenes; sticky studio correlation if uncaptioned; cutout look.
• Only lifestyle failure: clutter leaks into the concept; soft or occluded edges → deformed shape.

Bias toward clear product-first shots for geometry, then add enough lifestyle cells that backgrounds and support surfaces vary and get captioned. There is no binding 70/30 rule in the sources we cite. Caption every background you do not want fused to the trigger—same rule as character backgrounds. If bake still shows up, diagnose with overfitting vs undertraining.

Caption rules for products

Same philosophy as the captioning guide: trigger first, caption what should stay controllable. Specialize for objects.

What the trigger should own

• Fixed geometry / silhouette / unique design of this SKU.
• Permanent material identity when it defines the product (always brushed aluminum)—unless you want material promptable across a product line.

What captions should name (when visible)

1. Rare trigger first + class noun (zvpack bottle, not the real brand name as trigger).
2. View / orientation (front, three-quarter, rear, top-down).
3. State (open, capped, laced, unfolded).
4. Support surface / background / nearby props.
5. Lighting / framing / scale.
6. Material appearance as seen in that shot when lighting changes how it reads (cream glaze, brushed steel, matte rubber outsole).
7. Readable brand / packaging text: only if you have rights. Prefer captioning that a logo is present and its placement; do not use a real brand string as the trigger token; review OCR-heavy auto-captions.

FLUX-style:
A three-quarter front view of zvpack ceramic bottle on a walnut desk, cream glaze, metal cap visible, daylight from the right.

Tag-style:
zvpack bottle, three-quarter view, cream glaze, metal cap, wooden table, daylight, product photo

Bad:
beautiful product photo, masterpiece, red sneaker

Good:
sksshoe sneaker, lateral view, white midsole, mesh upper, concrete floor, soft daylight, full product shot

Illustrated anime props on Pony / Illustrious may prefer cleaned Danbooru tags—see the anime tagging guide.

Curation checklist: shape consistency (“object drift”)

Same curation mindset as identity drift, different tells. Object drift is shape / material / proportion inconsistency across the set or across seeds after training—not face morph.

1. Proportions stay stable (bottle height/width, sole thickness, blade length).
2. Materials do not flip (matte ↔ chrome, wood ↔ plastic) unless captioned as a lighting effect.
3. Logo/label layout is stable when text fidelity matters (no gibberish mutation).
4. No extra or missing parts (straps, buttons, screws).
5. No near-duplicates (same angle + same background).
6. Product is large enough in frame; no motion blur; silhouette not crushed by shadow.
7. One SKU only—no mixed colorways or product generations.

Export toward Kohya and Civitai

Save paired image.png + image.txt (same stem) with trigger-first captions. Nest for Kohya or zip flat pairs for Civitai using packing a LoRA dataset for Kohya and Civitai.

Ecommerce, IP, and brand consent

Doing this in LoRA Dataset

1. 01

   Upload one product reference

   Use a sharp shot that already shows silhouette and materials clearly.
2. 02

   Expand across the grid

   Generate object-consistent variations that rotate angle, lighting, and background while holding SKU geometry fixed.
3. 03

   Caption and download pairs

   Export paired image + .txt files, curate for shape consistency, then pack for your trainer. Free with your own Gemini API key; Google bills model usage at their rates.

LoRA Dataset does not guarantee perfect logo OCR, ecommerce conversion, or a fixed magic output count. It turns one reference into a captioned, object-consistent backbone you still curate.