03Finding the usabilityboundary of AIproduct images
The cross-border e-commerce team needed large volumes of furniture scene images for Amazon/Wayfair listings. Traditional shoots cost thousands per set. I used AIGC tools to explore a core question — which AI-generated images can ship directly, which need editing, and which must be regenerated.
Role
Product Intern · Cross-border E-commerce Product Management
Responsible for exploring and optimizing AI-generated product images for dining chairs, bar chairs, and single chairs. Using Flowith AI, Dreamer (Nano Banana), and Google Gemini for scene image generation — combining Amazon/Wayfair channel requirements to deliver 120+ scene cases and 50+ image tasks.
Problem
AI-generated images look good ≠ ready for e-commerce listings
Furniture scene images for overseas platforms need to show accurate product details, material texture, proper scale, and visual credibility. AI tools generate drafts in seconds, but outputs frequently fail on the details that matter most for actual product listings.
01
Color drift
AI frequently altered product colors. Work required repeatedly emphasizing "chair color 100% accurate" and "strictly match the reference image" — but models still drifted.
02
Angle & orientation loss
Needed precise control over product orientation — "rotate 90° to face front", "no bird's-eye view", "45° oblique downward" — but models interpreted spatial instructions inconsistently.
03
Quantity & placement errors
Specifying "one table, four chairs" might yield three or five. When replacing products in existing scenes, position and scale often shifted unpredictably.
04
CN→EN prompt gap
Chinese marketing language couldn't be fed directly to models. Translating into "instructions the model understands" was a separate skill — not just language translation, but concept translation.
Framework Design
6 scene templates × 5 operation modes: a reusable prompt production system
I decomposed the work into two dimensions — scene templates define "what to generate", operation modes define "how to generate and adjust". This meant each new task could start from an existing template library instead of writing prompts from scratch.
“The key shift was treating prompt work as a repeatable production method — not "think of a good prompt every time", but select from template → assemble → fine-tune → accept/reject.”
Dining Room
One table, four chairs, group dining scene — family of four or friends gathering
Bedroom
Vanity corner with female model — skincare or getting-ready scene
Reading Corner
Floor-to-ceiling bookshelf, female model reading, cozy atmosphere
Home Office
Male model working at desk with laptop, warm natural light
Cafe
Large cafe scene, three table-chair pairs, two people seated drinking coffee
Patio
American-style yard, large parasol + outdoor table + two dining chairs
Operation Modes
5 operation types that cover the full generation workflow
01
Scene Generation
Product white-background image + scene description → direct image generation. The most common starting point.
02
Reverse Prompt
Extract prompts from existing good images, then reuse for new products. "Reverse-engineer this scene photo's AI prompt for me."
03
Scene Replacement
Swap new product into existing scene composition. "Replace the bar chairs in image 4 with these 3 reference images, keep positions unchanged."
04
Attribute Fine-tuning
Precise adjustments to color, angle, perspective, and element additions. "Rotate the front chair 90°", "Change to 45° oblique downward", "Add tableware matching chair count."
05
Style Iteration
Keep product unchanged, experiment with different scene styles and atmospheres. "Don't add too many elements, let the image breathe. Try different styles."
Process
From product white-background photo to platform-ready asset
01
Requirement Decomposition
Identify product category (dining/bar/single chair), target scene, and channel requirements (Amazon main image vs scene image vs video first frame).
02
Template Selection & Assembly
Pick base framework from 6 scene templates, fill in product description, material, color, angle requirements. Add style constraints and negative prompts.
03
Multi-tool Trial
Dreamer is fast for batch iteration; Gemini has stronger comprehension for complex scene descriptions. Choose tool based on task type.
04
Iterative Refinement (3-5 rounds)
Address color drift, angle errors, and scale distortion by adjusting prompt segments. Operations include replacement, attribute tuning, and style iteration.
05
Quality Judgment & Delivery
Classify output into three tiers: ✅ ready for listing / ⚠️ needs minor post-editing / ❌ unusable, must regenerate.
Execution · Real Prompt Cases
From vague brief to precise visual instruction
Below are real prompts from actual work, showing how a vague request like "shoot a set of scene photos for this dining chair" becomes precise instructions an AI model can execute.
Dining Room (with models)
Commercial photography of modern dining chairs, one table four chairs, chairs matching the reference image. Set in a sunlit, warm beige dining room. 4 happy friends (American, ~30 years old, 2M 2F) sitting around a rectangular wooden table chatting and laughing. Soft natural light through large windows. Light wood floor, rug under table, decorative cabinet in background. Nordic interior, photorealistic, 8K, sharp focus.
Bedroom vanity (with model)
An elegant woman in her thirties wearing a champagne silk robe, sitting on the modern sculptural chair from the reference image, applying makeup with a brush at a bedroom vanity. Side-sitting pose, graceful manner. Minimalist vanity, large round frameless mirror on the wall. Carpet on floor, a large-leaf plant on the left. Warm afternoon sunlight from the side. High-end home magazine quality, cinematic lighting, minimalist style, 8K HD.
Scene replacement instruction
The first 3 images are white-background photos of the dining chair (same chair, multiple angles). Replace the dining chairs in reference image 4 with this chair, keep positions unchanged. Dining chair color accurate, 8K, realistic, front light.
Outcomes
What this practice demonstrated
AIGC production is a workflow problem, not an inspiration problem
Good outputs relied on repeatable templates + operation modes + quality gates, not on finding one "perfect prompt".
E-commerce usability is far stricter than looking good
Color off by a shade, angle slightly wrong, scale distorted = cannot ship. Aesthetic judgment must align with commercial standards.
Reusable templates cut trial-and-error costs in half
After establishing 6 scene templates + 5 operation modes, new product image generation went from "start from zero every time" to "pick template → fine-tune".
~25% conversion improvement vs traditional workflow
AI visual solutions demonstrated clear advantages in cost and speed over traditional photography, translating into higher content production efficiency and listing performance.
120+
Scene cases covered
50+
Image tasks completed
3-5
Iteration rounds per task
10+
Templates & SOPs delivered
Reflection