Date: 2026-08-05 · Status: Proposal pending final decision
Data sources: Grok deep research (under adversarial verification; pricing details pending) + dedicated scan of 8 competitor categories (57 entries, hallucinated items cleaned) + technical report from sibling conversation (unverified; overly optimistic figures corrected)
The competitor matrix is for directional reference only. Actual pricing/features should be checked against official websites.

1. Restating the Need

Build a small-business tool: end users—the customers of furniture stores—provide photos or videos of their homes plus product images of the furniture they want to buy, such as beds or sofas. The tool then outputs an overall styling preview of “this furniture placed inside the customer’s home,” helping answer whether the style matches and whether the item will fit.

The users are furniture-store sales associates or renovation-company staff. They need to demo it to customers on site and send the final image or video to the customer via WeChat—with video preferred.

Open decision: should this be a website, WeChat Mini Program, app, or desktop software?

2. Core Conclusion: Read This First

Platform: start with a mobile web page, i.e. H5. After validating paid demand, wrap it as a WeChat Mini Program. Do not build an app or desktop software.

In one sentence: the life-or-death point of this business is “a sales associate can produce a deal-closing asset in 15 minutes, demo it on the spot, and send the finished video via WeChat.” H5 requires zero installation, zero app-store review, can go live in a week, and works on any phone. Add a Mini Program only after money comes in, because it involves enterprise-entity category review and generative-AI compliance processes, which will slow validation down.

Technical route: 2D photo compositing + image-to-video. Do not touch 3D modeling or AR.

Specifically, it is “two images in, one video out”: room photo + furniture product image → cutout compositing, where AI only blends lighting and shadows without redrawing the furniture itself → preview image → feed it into an image-to-video model to produce a 5–15 second showcase video. This is the lowest-cost, fastest-to-launch path with the most “real-looking” output. Both 3D reconstruction—the Kejiale route—and real-time AR preview—the IKEA route—die on the same issue: every piece of furniture needs a 3D model. Furniture merchants simply do not have those. Modeling one item costs ¥100–500, and a one-person company cannot afford to build a model library.

Size judgment: do not attempt precise measurement. Provide a “three-level conclusion.” Ask the user to input one known dimension, such as “this wall is 3.2 meters” or a number from the floor plan. The system then converts the furniture footprint ratio and gives three levels: “spacious / just right / won’t fit,” plus footprint guide lines. Clearly label it as “for reference only; on-site remeasurement recommended.” Absolute measurement from a single photo is physically unreliable. Consumer products among competitors that claim to measure are also widely criticized by users for inaccuracy.

3. Competitor Landscape: Cleaned Version

Four Major Schools

1. 3D design tools: Kejiale (market leader, hundreds of thousands of renders per day, ¥29.8–166/month), 3D Home, Alibaba Everyflat/Tanping.
Common traits: first build a 3D floor plan, then place furniture models. The results are professional, but the learning curve is high and a sales associate cannot handle it in 15 minutes. Video export requires a pro plan. These tools serve designers, not sales associates.

2. AR placement previews: IKEA Place/Kreativ, Amazon View in Room, Wayfair, Home Depot.
Common traits: place furniture in real time through the phone camera. The on-site feel is strong, but they cannot output a finished asset that can be sent away—screen recording is the only workaround. Accuracy is criticized, and because domestic Android phones do not have Google ARCore, the experience is fragmented. IKEA Kreativ can only place IKEA’s own furniture.

3. AI photo makeover tools: RoomGPT, REimagineHome, Collov, Interior AI, SofaBrain, and many others.
Common traits: upload a room photo and get a redesigned room image. This is closest to what we want to build, but almost all of them focus on “whole-room style replacement.” They cannot faithfully place a specific product, do not handle dimensions, and do not output video.

4. General-purpose AI retouching tools: Doubao/Jimeng Photoshop, Meitu, Gemini (nano-banana), Qwen-Image-Edit.
Common traits: free or cheap. They can indeed “Photoshop a sofa into a living room,” but the furniture often gets altered—wrong texture, logo, or proportion—which is fatal when a sales associate is selling a specific product. They also do not provide sizing or video.

Four-Capability Comparison: Upload Real Photo / Place Specific Furniture / Judge Size / Output Video

• The only products covering all four are: Kejiale and 3D Home—both require pro subscriptions and 3D floor-plan modeling—plus Wayfair, Home Depot, which are internal tools for US retailers, and MeltFlex AI, a niche product.
• There is no dedicated domestic product for this lightweight combination: “one photo + one product image → faithful preview image + sizing hint + finished video.”
• A B2B tool for offline furniture-store sales associates—on-site demo + finished video sent via WeChat—is a clear gap.

What to Copy and What to Avoid

• Must copy: ask users for 1 reference dimension instead of requiring full measurement—the lowest possible friction; add a branded watermark to the finished video—sales associates will feel it is an official asset from their store, and it helps you spread; allow unlimited regeneration—customers keep changing it until they are satisfied, then pay.
• Avoid: locking the tool to your own product category—IKEA Kreativ can only place IKEA furniture; going pure AR; high per-use pricing—RoomGPT’s $1–2/use does not retain users; pure 2D sticker compositing without lighting/shadow blending—it looks fake and makes closing deals hard.

4. Technical Plan: P1 MVP

Pipeline: Two Images In, One Video Out

1
2
3
4
5
6

Customer's room photo ──┐
                        ├→ ① Furniture cutout → ② Perspective placement + AI lighting fusion → ③ Preview image
Furniture product image ─┘                                          │
One known dimension (user input) → ④ Footprint ratio calculation → three-level conclusion + guide lines │
                                                                    ↓
                                          ⑤ Image-to-video (pan/track showcase 5-15s) → ⑥ Watermarked MP4 → send to customer via WeChat

① Cutout: BiRefNet / rembg (open source, local execution, free). Most product images have white backgrounds, so cutout quality is high.

② Compositing (the core factor determining whether it “looks real”): prepare two routes—

• Main route: perspective transform placement (homography) + soft shadow, then use an AI image-editing model for furniture-region-protected lighting/shadow blending. Only redraw the lighting and shadows around the furniture and on the contact surface; the furniture itself remains pixel-perfect. This prevents the product from being distorted.
• Model channels: domestically, use Tongyi Qwen-Image-Edit (Bailian key already available, direct connection without proxy); backup option: Gemini 2.5 Flash Image (nano-banana, AIStudio2API 7870-7872 already deployed and testable).

③ Three-level sizing: user inputs 1 known dimension, such as wall width, ceiling height, or a floor-plan annotation → calculate pixel scale → convert the furniture dimensions—length, width, height from product details, either manually entered by the sales associate or parsed from a link—into footprint → output “spacious / just right / won’t fit” + draw footprint guide lines on the preview image. Optional enhancement: use Depth Anything V2 for relative depth to make occlusion more realistic, e.g. the furniture being partially blocked by a coffee table.

⑤ Video: wan2.6-i2v-flash (Bailian, full pipeline already tested end to end: async submit → polling → MP4 download). 5s/720P/silent = ¥0.75; 15s ≈ ¥2.25. Prompting uses “slow camera push-in / orbit showcase of the living room.” The calling code from the motion-comic pipeline can be reused directly.

⑥ Delivery: MP4 with watermark—store name/logo + “AI-generated effect for reference only,” with the watermark also serving compliance requirements → sales associate saves it to the phone album → sends it to the customer via WeChat. Also generate an H5 finished-video page link that can be directly forwarded and opened in WeChat.

Platform and Deployment

• Frontend: mobile H5 (Astro or Next.js), deployed on Cloudflare Pages (existing infrastructure, zero additional cost).
• Backend: Cloudflare Workers + R2 for image/video storage; or reuse the existing Docker host. Video generation runs as async tasks—Bailian already uses an async polling model.
• Sales-associate accounts: P1 can skip an account system; the URL can carry the order number. Add accounts in P2.

5. Compliance Red Lines

As of 2025-09-01, the “Measures for Labeling AI-Generated and Synthetic Content” are in effect and must be followed.

1. Preview images and videos must carry an explicit “AI-generated” label. The watermark conveniently serves two purposes.
2. Use API models that have already been filed/registered, such as Tongyi and Kelvin. We do not need a separate filing ourselves, but the labeling obligation is ours.
3. Product-image copyright: in the sales-associate scenario, merchants use their own product images, which is fine. Do not crawl other stores’ product images to build a library.
4. WeChat Mini Program category review only applies in P3. The H5 stage does not involve it.

6. Implementation Plan

P0: Service-Style Validation (3–5 Days, No Code, ~¥50 API Cost)

Goal: first prove someone is willing to pay, then write a single line of code.

1. Manually run the pipeline: find 3 real scenarios—your own home, a friend’s home, or online buyer-show photos—plus real Taobao product images. Use Bailian Qwen-Image-Edit + wan2.6-i2v-flash to produce one 15-second finished video for each.
2. Take the 3 samples to 5–10 furniture stores / building-material market stalls—or 1688 furniture merchants. Demo it and ask: “¥19.9 per video. Your customers use it while viewing furniture. Would you buy it?”
3. If ≥2 give a clear willingness to pay → proceed to P1. If not → stop the project here, losing only 3 days + ¥50.

Note on project focus discipline—one in, one out: if this project gets a GO, decide which current project gives way, or clearly define it as a derivative scenario of Lynxhouse, i.e. adjacent monetization between real estate and furniture.

P1: H5 MVP (1–2 Weeks)

• Upload two images → cutout → compositing → three-level sizing → video generation → watermarked MP4 download / finished-page sharing.
• Manual QA fallback: a human reviews the generated result before sending it to the customer (service-style DNA; protect reputation early on).
• Payment: offline WeChat payment is fine at first. Do not delay launch because of payment integration.
• Validation goal: 10 paid orders.

P2: Polish (2–4 Weeks)

• Bulk import furniture-store catalog, i.e. the store’s in-stock product list, so sales associates select products instead of uploading.
• Custom branded watermark, sales-associate accounts, order management, WeChat Pay merchant account.
• Pricing reference: ¥19.9 per video; ¥99/month for one sales associate; ¥499/month for multi-account furniture stores. Anchor: Kejiale pro version at ¥166/month.

P3: Expand as Needed — All “After Making Money”

• WeChat Mini Program (requires enterprise entity and tool category; prepare AIGC labeling description in advance).
• Multi-angle previews: composite multiple photos of the same room and cut them into a carousel video.
• AR on-site demo, using Mini Program AR plugins such as Kivicube—only build it if customers explicitly ask for it.
• High-end 3DGS room walkthrough version, using Luma/Polycam capture + gsplat—only for premium individual orders.

7. Cost Estimate

(Exact API pricing should be supplemented by the Grok-verified report, but it does not affect the architecture decision.)

8. Risk Table

9. Complexity Assessment

• P0: Low (manual API calls, 3–5 days)
• P1: Medium (H5 + orchestration of three API stages + watermarking, 1–2 weeks)
• Full project through P2: Medium. Solo-buildable. No new framework required—Astro/CF/Bailian pipeline are all already available.

Appendix (added 2026-08-06, completed): adversarial verification results are in ~/furniture-ar-viz-verify-appendix-2026-08-06.md — 74 “rescue votes” confirmed and 7 claims debunked (Flux pricing reversed / Kontext free commercial use false / ARKit plug-and-play fantasy / AR conversion rate of 25–40% without source / NYU 0.119 outdated / specific 3DGS figures unsourced / Tencent Arrow Lab misattribution); the 08-06 restored workflow finished running (105 agents, 25 claims → 4 survived + 21 killed, 5 comprehensive conclusions all validate the route in this report). New debunks: Flux Kontext actual price is $0.04/image; precise scale from pure SLAM / single-image depth / WebAR is all overhyped. Core conclusion is unaffected; if anything, it confirms that delaying AR is the right call.