What the small workshop changed

Tischlerei Dein Freund is a German joinery with 10–49 employees that designs custom furniture and interiors. Early in a project, a customer needs more than one polished rendering. They need to compare several understandable directions for form, proportion, material, and the character of the future object.

Before the project, the company used Midjourney. According to the European Digital Innovation Hubs Network case page, one satisfactory image required 30–50 requests, with each request taking 30 seconds to one minute. Visual exploration became a waiting exercise and made it difficult to revise a concept with the customer in real time.

Through the Test Before Invest program, DIH4AISec deployed a local Stable Diffusion installation with the AUTOMATIC1111 interface and trained the staff. The system could accept simple sketches and produce a variant in roughly three seconds. The case authors estimate time savings of more than 95% in the best cases.

The result needs careful interpretation. The public page does not disclose the graphics card, resolution, sampling steps, exact checkpoint, or methodology for deciding when an image is satisfactory. It also does not calculate financial ROI: it lists €20,000 of public support and no direct investment from the joinery. The 95% figure is therefore not a promise for every company. It shows that one narrow local workflow became fast enough to use during a consultation.

The business task is not image production

Custom furniture is sold by reducing uncertainty. Customers struggle to imagine an object from words, while a complete 3D model is too expensive for every early idea. The employee must choose between a slow, precise visualization and quick but weak communication.

A generative model fills the middle layer. It helps the parties discuss style and composition before an engineer spends hours on dimensions, joints, and a bill of materials. The gain appears when an image shortens the question-and-answer cycle, not when it merely increases the number of attractive files.

An AI concept must not be presented as an engineering design or an exact image of the finished product. The model may change the number of sections, panel thickness, wood-grain direction, fittings, or geometry between variants. Feasibility, price, and schedule remain the craftsperson’s decisions, while the contract must rely on an approved drawing and specification.

A reproducible workflow

A practical process has six steps.

1. **Brief.** Record product purpose, dimensional range, style, materials, budget, and room constraints.
2. **Reference pack.** Add a sketch, permitted wood samples, room photographs, and detail examples that the business has the right to use.
3. **Controlled generation.** A template fixes viewpoint, product type, and invariants. A sketch or contour holds composition more reliably than a free-form text prompt.
4. **Human selection.** The designer removes infeasible variants and shows the customer a small selection rather than a stream of dozens.
5. **Decision record.** The chosen concept receives an ID, model version, generation parameters, and a link to the project record.
6. **Engineering handoff.** An engineer creates the CAD model, material list, and price; the AI image remains a visual reference.

What local deployment adds

In this case, local deployment was not an abstract security requirement. It contributed directly to performance and control. Images were generated on company hardware without a remote-service queue, while employees used a familiar interface and simple sketches.

For a Russian small business, local generation is especially relevant when the inputs contain:

  • photographs of homes, offices, and private interiors;
  • floor plans and address-related information;
  • products that have not been announced publicly;
  • proprietary shapes, collections, and materials;
  • customer data tied to an order;
  • a high iteration volume that makes cloud variable cost material.

The minimum environment includes a dedicated workstation, local project storage, a registry of approved models and extensions, configuration backups, and role-based access. The web interface should not be exposed to the internet unless there is a specific need.

AUTOMATIC1111 documentation warns that installing extensions is equivalent to permitting arbitrary code execution. The interface blocks those operations under remote-access configurations, and its maintainers advise against enabling the insecure bypass for normal use. A workshop therefore needs a simple rule: one responsible employee installs reviewed extensions, and the production workstation does not become a shared experimentation machine.

The interface license and model license are separate

Running locally does not make every downloaded checkpoint free for commercial use. The operating record should preserve the exact model name, file source, hash, download date, and license text for that version.

AUTOMATIC1111 is distributed under AGPL-3.0. Model weights have separate terms. Stability AI’s current licensing page says its Core Models are free for commercial use by organizations with annual revenue below USD 1 million, while different terms apply above that threshold. This rule cannot be assumed for every file carrying the Stable Diffusion name: third-party fine-tunes, LoRAs, and other generations of the model may have separate restrictions.

Before a pilot, answer four questions:

  • does the exact weight version allow commercial use;
  • may the output be shown and transferred to a customer;
  • where did the references come from, and are they licensed for this use;
  • who checks similarity to third-party trademarks, designs, and characters?

Legal review concerns more than the model. A customer-provided reference may itself reproduce a protected design, and the generation may follow it too closely.

Data for consistent quality

The workshop’s durable asset is not a collection of random prompts, but a library of reviewed decisions. For every useful result, preserve:

  • a minimized brief;
  • the source sketch and permitted references;
  • product family and material;
  • model parameters and a fixed seed when reproduction is required;
  • the selected result and the reason for selection;
  • an engineer’s feasibility notes;
  • a photograph of the finished item if the order was completed and the customer consented.

This chain shows which visual promises fail in production. If the model draws supports that are too thin or hides an impossible joint, add a review step or template constraint instead of endlessly rewriting prompts.

Do not train a custom model on every project immediately. Reproducible templates, an approved-material catalog, and controlled sketch transformation are enough at first. Fine-tuning makes sense after the business has a clean dataset and a measurable improvement target.

Measuring the effect

The EDIH case compares generation speed, but management needs metrics for the complete process:

  • minutes from brief to first usable concept;
  • generations required before selection;
  • consultation duration;
  • revision cycles before CAD;
  • share of concepts rejected by engineering;
  • discrepancies between visualization and approved specification;
  • consultation-to-paid-project conversion;
  • employee time spent maintaining the local installation.

A fast renderer may save nothing if the customer receives twenty variants and takes longer to choose. The interface should therefore limit the selection, while the employee explains which parameters are not yet fixed.

Illustrative economics for a small business

The €20,000 in the German project funded implementation and training through EDIH. It is not the market price of a local deployment and was not a direct cash investment by the joinery. A business budget should separate the workstation or reuse of existing hardware, setup, training, upgrades, backups, and administrator time.

Consider a model with 40 consultations per month, 20 designer minutes saved per consultation, and a fully loaded labor cost equivalent to RUB 1,800 per hour. That releases RUB 24,000 of capacity per month. If monthly support and depreciation exceed this amount, the project must pay back through higher conversion or additional order volume, measured separately.

This is an illustrative model, not a result from the source case. Do not count time reduction, additional sales, and an avoided subscription as three independent benefits unless the measurements prove they do not overlap.

A pilot on one product type

The experiment can start in four weeks.

  • **Week 1:** select one category, such as cabinets, and assemble 20–30 permitted briefs and sketches.
  • **Week 2:** configure the local workstation, freeze the model version, and build two or three generation templates.
  • **Week 3:** run old and new consultation flows in parallel without presenting an AI concept as a drawing.
  • **Week 4:** compare time to decision, revision count, engineering feedback, and customer response.

The success criterion is not three seconds per file. The pilot succeeds if customers choose a direction sooner, engineers reject fewer visual promises, data stays inside the approved environment, and employees can reproduce results after a system update.

The management takeaway

The joinery case has a useful scale for local AI: one clear stage, frequent iterations, expensive waiting, and a mandatory human decision. Technology strengthens the consultation but does not replace engineering documentation.

Before buying hardware, select one product category and measure the complete path from brief to approved concept. Record the model and license, restrict extensions, separate AI imagery from CAD, and count total cycle time. If only image output became faster while approval stayed unchanged, the wrong bottleneck was automated.