AI for furniture makers: visuals of custom pieces, offers from sketches and EUDR records

What AI does for a furniture maker

A furniture, kitchen or joinery maker can use AI in three places: pictures of a custom piece in the customer's own room, offers drafted from sketches and measurement sheets, and the timber records that the EU Deforestation Regulation (EUDR) asks for. One NVIDIA RTX PRO 6000 Blackwell Workstation Edition or Max-Q card with 96 GB of memory runs the image work in a workstation in the design office; a server with two Server Edition cards adds offers, product texts and EUDR records for the whole company. Room photos, drawings and prices stay on hardware you control.

The image model starts from the drawing in your configurator or CAD system, and the offer text is written around the price your system calculates; the model never sets a price.

We wrote this for owners, managing directors and heads of sales or design at furniture, kitchen and shopfitting companies without a large IT department. Technical terms are explained where they first appear, and the hardware is in one section with a table. Agencies that make campaign images have different needs, covered in our guide to an AI server for creative agencies; ray-traced rendering of production scenes is the subject of our GPU render server guide.

Furniture and kitchen visualisation with AI, before the piece is built

A custom wardrobe or a kitchen is usually planned in CAD or in a configurator, and the customer sees a technical rendering, often in neutral surroundings. An image-editing model can take that rendering and a photo of the customer's room and combine them into one picture: the wardrobe in the alcove, the kitchen along the real wall, under the light of the real window. It can then change fronts, handles, worktops and colours on request, so the sales conversation works with pictures instead of material samples alone.

Two models are under licences that allow this commercial use. Qwen-Image-Edit-2511 is a 20B model stored in BF16, a 16-bit number format. Its card lists "enhanced industrial design generation", "material replacement for industrial components", better consistency between edits and support for LoRA add-ons, which are small trained extensions, for example for your own surfaces. Its example passes two input images to one pipeline, which fits the pair of room photo and product rendering. FLUX.2 [klein] 4B, with 4B parameters, is smaller: its card says it fits in about 13 GB of GPU memory, handles text-to-image and multi-reference editing in one model, and its example runs four steps, so it suits quick variants of colour and fronts.

Hugging Face model cards, file lists and licence files, read 25 October 2026. The Qwen-Image-Edit-2511 repository has no separate licence file; its card states Apache 2.0.

Other well-known image models carry non-commercial licences, as our guide to FLUX GPU requirements lists, so read the licence file of every model before it goes into sales work.

A generated picture is an impression of the piece, and it can be wrong in ways a customer will remember: a drawer too many, a handle in the wrong place, proportions that shift between two edits. Dimensions, fittings and the cut list come from the CAD drawing. Before a visual goes into an offer, someone compares it with the drawing, and the offer marks it as an AI visualisation. Article 50 of the AI Act, which has applied since 2 August 2026, makes deployers disclose deep fakes, as our article on the EU AI Act explains; whether a room visual falls under that is for your legal counsel, and a clear label costs nothing.

We found no published speed figure for Qwen-Image-Edit-2511 on RTX PRO cards. For the related Qwen-Image generator, our guide to a Qwen-Image server estimates at most about 216 images an hour at 50 steps on one RTX PRO 6000 Server Edition, an upper bound, and an edit costs more because it also processes the input pictures. Even a tenth of that bound would be about 21 pictures an hour, enough for several offers a day; measure your own edits on one card before you plan more.

Room photos from private customers show their homes, sometimes people and family pictures, so they are personal data under the GDPR. On your own server they stay under your access rules, and your data-protection officer sets how long they are kept.

Custom furniture quotes from sketches, measurements and room photos

Many custom orders start on paper. A customer brings a sketch with dimensions, a fitter measures the room, and photos show where the window, the radiator and the sockets are. Someone in the office then types all of it into the configurator before a price exists. A vision-language model, a language model that also reads images, can do the first pass. Qwen3.8-27B is described on its card as "a native vision-language model that understands images and videos"; the card does not mention handwriting or sketches, so test it with twenty of your own before you rely on it.

The model turns the sketch and the measurement sheet into a structured list: wall lengths and heights, positions of windows, doors, pipes and sockets, the pieces the customer asks for, and for each value whether it was read from the sketch or is missing. A person checks the list against the measurement and enters or confirms it in the configurator. The configurator, or the ERP behind it, calculates the price, the parts list and the delivery terms. The language model then writes the offer text: a description of each piece in plain words, the options with their differences, and a list of open questions for the customer, such as a socket position the sketch does not show. The price is never written by the model. Our guide to AI offer writing for machine and plant builders explains that division in detail, and orders that later arrive by email are covered in our guide to AI order entry into the ERP.

We supply the RTX PRO 6000 Workstation Edition, Max-Q and Server Edition as cards, and workstations and servers built around them; send us a typical sketch, the configurator you use and the number of offers a week, and our team replies with a configuration.

EUDR due diligence records for wooden furniture

Regulation (EU) 2023/1115, the EUDR, covers wood among seven commodities, and the Commission's page names furniture among the derived products. Annex I lists "9403 30, 9403 40, 9403 50, 9403 60 and 9403 91 Wooden furniture, and parts thereof", and seats of wood under heading 9401. After the amendments of December 2024 and December 2025, Regulations (EU) 2024/3234 and 2025/2650, the obligations apply from 30 December 2026. For micro and small operators the date is 30 June 2027, as the Commission's page sets out, unless they were already covered by the EU Timber Regulation.

What a company has to do depends on its role. An operator that places relevant products on the EU market or exports them, for example by importing timber or furniture from outside the EU, submits a due diligence statement before doing so (Article 4(2)), and collects for each product, among other information, the geolocation of the plots of land where the commodity was produced, the date or time range of production, the quantity and the supplier (Article 9(1)). The 2025 amendment added the downstream operator: a company that places on the market products made from relevant products, all of which are already covered by a due diligence statement or a simplified declaration. A maker that builds wooden furniture from boards bought in the EU may be one. Under Article 5(3) it keeps the name, address, email and, if available, web address of its suppliers, the reference numbers of their due diligence statements where the supplier is an operator, and the same contact details of the downstream operators or traders it supplies. It keeps this for at least five years and provides it to the competent authorities on request (Article 5(4)); non-SME downstream operators also register in the EU information system (Article 5(2)). Which role your company has, and which duties follow, your legal counsel confirms. This is general information, not legal advice.

The paper side is where a language model helps. Delivery notes and invoices for boards, veneers and solid wood arrive as PDFs in each supplier's own layout. The model reads the supplier's details and the reference numbers into a register linked to your purchase orders, marks deliveries without a reference number for a query to the supplier, and assembles the record when an authority or a customer asks. The register itself is a database that people confirm; the model fills it and checks it. Our guides to supplier declarations with AI and to the digital product passport describe the same pattern for other product rules.

Which hardware, in plain words

Two numbers decide the hardware. A model has to fit into the memory of the graphics card, the GPU, and the work in progress needs room beside it. Our users-per-card rule takes 90 per cent of the 95.6 GiB that the driver reports for a 96 GB card, less 3 GiB per card: 83.0 GiB for weights and work.

Memory by our users-per-card rule from the file sizes on Hugging Face (read 25 October 2026); power at the wall by our estimate from card and processor power plus 15 per cent for the rest, as in our workstation and server guides, confirmed by a measurement on the delivered system.

A workstation in the design office. The edit model's 57.7 GB of files are 53.7 GiB, which fits on one 96 GB card in 16-bit with room left to work, and is more than a 48 GB card holds. A workstation with one RTX PRO 6000 Workstation Edition, the full 600 W card, draws about 1.1 kW at the wall by our estimate, as our page to buy an RTX PRO 6000 Workstation Edition PC works out, and runs from an ordinary 230 V office circuit. The RTX PRO 6000 Max-Q is the same chip at 300 W with a blower cooler, and NVIDIA allows up to four in one workstation, so a company that expects to grow can start with one Max-Q. The sketch reader and the edit model take turns on one card, because both together leave too little room to work by our rule.

A server for the whole company. The entry server has two RTX PRO 6000 Server Edition cards, two AMD EPYC processors, 384 GB of DDR5 ECC and two NVMe drives. Card 1 runs gpt-oss-120b for offer texts, product descriptions for dealers and the EUDR register; by our rule it answers 19 requests of 32K tokens at the same moment, where a token is a piece of a word and 32K tokens hold a long offer with its background. A request counts only while it is being answered, so many more people can use it. Card 2 holds Qwen3.8-27B in FP8 and FLUX.2 [klein] 4B together, 28.75 and about 12 GiB, which leaves about 42 GiB of the 83.0 GiB for work by our arithmetic. The edit model stays on the workstation in the design office, or the server gets a third card. The server draws about 1.8 kW and belongs in a server room; without one, our Private AI/ML service runs the same platform on dedicated hardware in a Tier-3 data centre in Lithuania, inside the EU.

For continuous operation in a rack we recommend the Server Edition; the Max-Q belongs in a rack only in a chassis whose maker lists it. Our article on AI servers for companies of 10 to 100 people shows how one server serves several departments.

Notes for the IT team

1. Image pipeline. The Qwen-Image-Edit-2511 card's example uses QwenImageEditPlusPipeline from Diffusers with two input images, 40 steps and true_cfg_scale 4.0; FLUX.2 [klein] 4B's example uses four steps. Keep the rendering from the configurator at a fixed size and record the prompt, seed and model revision with each visual, so an approved picture can be reproduced.
2. Photos as tokens. Qwen3.8-27B's preprocessor uses patches of 16 pixels merged two by two, one image token per 32 × 32 pixels by our reading, and caps an image at 16,777,216 pixels. A 12-megapixel phone photo is therefore about 11,700 tokens by our arithmetic, and with 64 KiB of 16-bit cache per token plus 144 MiB of fixed state it takes about 0.9 GiB per request. Scale photos to the detail the sketch needs.
3. Configurator and ERP. Give the platform read access to the configurator's article data and fetch prices only through a defined interface into the offer template; the model writes nothing into the ERP.
4. Engines and network. vLLM serves gpt-oss-120b and Qwen3.8-27B through an OpenAI-compatible API behind an LLM gateway with keys and a query log. vLLM's --api-key protects only paths under /v1, /v2, /inference and /cohere (vLLM 0.32.0 security guide), so keep the engines on an internal network. Run a current vLLM release (0.32.0 of 18 October 2026 at the time of writing), the NVIDIA Container Toolkit 1.17.8 or later and a driver release that NVIDIA lists as fixed in its bulletin of 30 September 2026.
5. Software on delivery. Workstations and servers arrive on request with the operating system, the NVIDIA driver, CUDA and a container runtime; our guide to RTX PRO Blackwell software support lists the versions for these cards.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We supply the RTX PRO 6000 Workstation Edition, Max-Q and Server Edition as cards, and build workstations and servers around them to order, assembled and tested before delivery, delivered anywhere in the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Our Private AI/ML service, with our engineering partner Vixen.UNO, sets up the models, the offer templates and the EUDR register with the connections to your configurator and ERP. To start small, describe one recent custom order, from the first sketch to the offer, and we propose a pilot and the hardware for it.

FAQ