90% cost cut — it's real
Numbers first: a Nordic-style solid wood furniture brand, roughly ¥30M annual revenue, used to spend ¥400K/year on product photography. After AI, it dropped to ¥40K.
Not bragging. But ¥30K of that ¥40K is still real photography — core bestsellers still get real shoots. AI only took 80% of the workload.
The furniture industry pain point is concrete: each SKU needs 5-8 photos — white background, lifestyle scene, detail shots. One set runs nearly ¥10K. 50 new SKUs = ¥500K. Save that, and you can fund a whole new product line.
What they used
Simple stack:
- Midjourney V8.1 — main image generation, scenes and ambiance
- ComfyUI — precise product image control, keeping SKU consistency
- Photoshop AI — post-production, background swap, detail fixes
Just 1 designer handling it. Not a big production. The key was getting the workflow right.
How the workflow runs
White background: ComfyUI takes over
White background shots are simplest. Photograph the product, cut it out, feed into a ComfyUI workflow, auto-generate standard white-background images. Used to take a photographer half a day; now 50 images in 20 minutes.
Scene images: Midjourney + real photo anchor
Scene images are the main event. Previously: rent a space, move furniture, set up lighting, shoot. One scene set = 2-3 days, ¥10-20K.
Now: shoot one white-background product photo → feed to Midjourney as reference → write prompt describing scene ("Nordic living room, floor-to-ceiling windows, afternoon sun, wooden floor") → generate 5-8, pick 1.
15 minutes to output, near-zero cost.
Detail shots: Photoshop AI fills in
Wood grain texture, hardware close-ups — real photography is still more reliable. But occasionally need to supplement angles; Photoshop AI's generative expand does the job.
3 mistakes — you need to hear these
Don't think it was smooth. Three big pitfalls.
Mistake 1: AI furniture "looked wrong"
At first, straight text-to-image with Midjourney. Prompt: "a Nordic oak dining chair." The output looked nice — but the leg curve, backrest proportions were completely off from the real product.
Clients received goods and complained: "Looks different from the photo."
Nearly triggered a return wave.
Fix: Switch to image-to-image. Shoot one real product photo, use as reference, prompt only describes scene and ambiance — not the product itself. Now the product form is accurate, the scene is AI-generated.
This is the critical step — the product must be real, the scene can be fictional. Get this backwards and it's a disaster.
Mistake 2: Inconsistent style
50 SKUs of scene images, generated in 3 batches. Result: 3 subtly different styles — lighting angles, color tones, even floor textures didn't match. Put them in one product catalog, looks like 3 different brands.
Fix: Build a style parameter template. Lock one set of Midjourney parameters (style reference + color palette + lighting description), shared across all SKUs. Designer saves it as a txt file, copy-paste every time.
Mistake 3: Client asks "is this AI?"
A B2B client looked at the product images and asked directly: "These are AI-generated, right?"
Awkward. Not that you can't admit it — but clients doubt product quality: "If even the photos are AI, is the actual stuff reliable?"
Fix: Core bestsellers keep real photography for main images and detail page headers. AI-generated images go to scene pairings, content marketing, social media. Two types, separate roles. Clients don't question it.
The ledger
- Photography costs: ¥400K/year → ¥40K/year (saved ¥360K)
- New product speed: 50 SKUs from 2 months → 2 weeks
- AI tool costs: Midjourney $30/mo + ComfyUI local (one-time RTX 4060 Ti) + PS AI $20/mo, ~¥4,000/year
- Headcount: No layoffs — photography budget redirected to designer salary
3 takeaways
1. Real product, fictional scene — use real photos as anchors, AI only generates the environment, never let AI invent the product
2. Lock style parameters — save as reusable template, don't re-tune every time, or SKUs will look inconsistent
3. Keep real photography for core products — AI images for marketing and long-tail SKUs, bestsellers still need real shoots. Client trust matters more than saving money.
In furniture, images are your storefront. AI can save big money, but cut costs in the wrong place and you'll damage the brand. Use AI where it belongs — scene images, content images, long-tail SKUs. Core bestsellers still get real photography.