How to Make 3D Product Visualization via Meshy Plugin (2026 Guide)

You can make 3D product visualizations in ChatGPT by installing the Meshy plugin, describing the product in plain words, and letting the agent generate the 3D model, texture color variants, and split it into named parts. Then ask it to build a web page where shoppers drag to rotate and watch the piece assemble. One armchair prompt in, a 20-piece furniture catalog out.

That is what the video above shows, and it is a real Meshy workflow, not a mockup. The agent driving it is GPT-6 Astra, OpenAI's agent model, running the Meshy plugin inside ChatGPT. You can see the Meshy task panel with the image-to-3D endpoint, the 30-credit cost, and the 1m46s generation time. You can see Auto Split return the chair as six separate parts, each tinted a different color in the preview so you can tell them apart. And you can see the finished site, with a slider that puts the chair together rail by rail. This guide turns that 54-second clip into steps you can copy.

What Is 3D Product Visualization?

3D product visualization means showing a product as an interactive 3D model instead of a flat photo. The viewer can rotate it, zoom in on a joint, switch the fabric color, or pull it apart to see how it fits together. It is the difference between a catalog picture and picking the chair up in your hands.

It is not the same thing as 3D product rendering. Rendering produces still images or video from a 3D model, usually for ads and listings. Visualization keeps the model live in the browser so the shopper does the exploring. You often want both, and they start from the same asset: a clean, textured 3D model of the product. If you are still designing that product rather than showing a finished one, Meshy's 3D product design workflow covers the sketch-to-prototype side. This guide picks up once the design is settled.

That model is the expensive part. Furniture 3D modeling used to mean a 3D artist for each piece, so a brand with 20 pieces paid for 20 models, plus a second round for every fabric option. The workflow below replaces that with a chat window. ChatGPT plans the collection and writes the web page. Meshy makes the geometry and textures. You approve things and pay in credits.

Can ChatGPT Do 3D Modeling?

Not by itself. ChatGPT cannot produce a mesh file, and neither can GPT-6 Astra, the agent model in the video. What it can do is run a 3D generator for you. With the Meshy plugin installed, you describe a product and the agent creates a Meshy task, waits for it, checks the result, and saves the file to your computer. It also handles the boring parts: naming files, tracking 20 jobs, and writing the site that displays them. If you have used Meshy's 3D Agent on the website, this is the same idea running inside ChatGPT.

If you only want the install steps and a first test model, read How to Use Meshy in ChatGPT & Codex. If you want the API-key route with Blender at the end, that is the GPT-6 Astra workflow guide. This page is about the product catalog job: many pieces, color options, named parts, and a page to show them on.

How to Build a 3D Furniture Catalog in ChatGPT with the Meshy Plugin (7 Steps)

The steps follow the video in order. Step 1 is setup. Steps 2 to 4 build and refine one hero product. Step 5 scales it to a collection. Steps 6 and 7 put it on a web page and ship it.

One thing to set straight before you start: the plugin runs Meshy tasks under the hood, and every task uses credits from your Meshy account, the same credits the Meshy website uses. The agent tells you the estimate before each job. A whole catalog is dozens of jobs, so the cost math in the later steps is not optional reading.

Step 1: Install the Meshy Plugin in ChatGPT

Open the ChatGPT desktop app, go to the Plugins page, and search for Meshy. The listing reads "Meshy: Generate 3D models." Click install, wait for it to finish, then start a new chat. Plugins only appear in chats you open after installing. In the chat, start your message with @Meshy so ChatGPT knows to use the plugin.

The first time you ask for a model, the agent shows a sign-in link and a short code. Open the link, log in to Meshy, approve, and you are done. Later chats reuse that sign-in. You never paste an API key into the chat, and if ChatGPT ever asks you to, say no and ask it to sign you in through the browser.

You need a paid Meshy plan. Free accounts cannot generate through the plugin. Models made this way are saved on your computer, not in your Meshy web workspace. If you want to keep editing one on the Meshy site, upload the file there. Check pricing before a long catalog session.

One more setting before you start. The plugin works with any model in the selector next to the message box, but in the video it is set to GPT-6 Astra, and that is what we used in our test too. Planning a 20-piece collection and writing a web page are multi-step agent jobs, which is what Astra is built for, so pick it if your plan offers it.

For a full walkthrough of the install, sign-in, and the most common setup errors, use the ChatGPT and Codex guide linked above. The rest of this tutorial assumes the plugin is working.

Step 2: Generate the Hero Product from a Prompt

Pick one piece to be the star. In the video it is an armchair, and this is the exact message:

@Meshy Generate a tall modern wooden armchair with continuous natural oak side frames integrating the legs and armrests. A curved walnut seat and broad walnut backrest, with two oak support rails beneath the seat. Rounded edges, slender proportions, matte wood finish.

Notice what the prompt does. It names the form (tall armchair), the materials by part (oak frames, walnut seat and backrest), the construction (side frames that flow into legs and armrests, two rails under the seat), the proportions (slender, rounded edges), and the finish (matte). Every one of those words shows up in the model. Vague prompts like "a nice wooden chair" give you a generic chair with guessed materials.

The agent checks your balance, reports the estimated cost, and runs the job. In the video the task is a textured image-to-3D run that cost 30 credits and took 1m46s. The image-to-3D API reference explains what the agent is calling. Your numbers will differ a little by model and settings. When it finishes, the agent saves a GLB file to your folder and shows a preview. Spin it around. If the rails are missing or the seat is flat, say so and ask for a new version before you spend credits on variants.

Step 3: Create Color and Material Variants with Retexture

Shoppers want options, and options are where the old workflow got expensive. With Retexture you keep the geometry and only change the surface. The video asks for "the chair with 3 different colors." Be more specific than that. Taken literally, that line describes one chair with three colors on it, which is a single Retexture job. What you want for a catalog is three separate versions:

@Meshy Make three retextured versions of the chair, one file each, same geometry: (1) olive green wool upholstery with a light oak frame, (2) cognac leather with a black-stained oak frame, (3) cream linen with a walnut frame.

Retexture is a separate Meshy task, and one task produces one new texture set, as the Retexture API reference spells out. Three versions means three tasks. At the published rate that is 10 credits each, so 30 credits for the set. The agent runs them one after another and saves each variant as its own file next to the original. It also checks the result against your request and tells you when a color did not land, which is your cue to rerun that one variant rather than all three.

Two habits save retries here. Say "versions" or "variants" and "one file each," so the agent does not merge your colors into one texture. And name materials, not paint colors: "olive green wool upholstery with a light oak frame" reads better to the model than "green." Because every variant sits on the same mesh, they swap cleanly in a web viewer later. That is the whole trick behind the color picker in Step 6.

Step 4: Split the Model into Named Parts with Auto Split

This is the step that makes the final page special. Instead of one solid chair, you get six pieces that can fly apart and snap back. The video asks for exactly that:

@Meshy Use Meshy's Auto Split API to separate the chair into left side frame, right side frame, seat, backrest, front rail, and rear rail.

Auto Split cuts a model along its natural part boundaries. It started as a 3D printing feature, and the same cuts work for an exploded view on a web page. Run on its own, in auto mode, it decides the parts. Given a list like the one above, it runs in by_parts mode and uses your names, up to ten of them, to steer the cuts. Each part comes back as a separate object inside one GLB, OBJ, FBX, USDZ, BLEND, or 3MF file. The task costs 10 credits. Details are in the Auto Split API reference.

Two habits get you the clean six-piece result from the video. First, name the mode: add "with mode by_parts and this exact prompt" to your message, so the agent runs a named split instead of the default auto mode, which cuts along every boundary it finds and gives you a finer breakdown than a product page needs. Second, match the names to how the chair is actually built. Auto Split follows the geometry, so a chair with two rails and two stretchers will come back with rails and stretchers as separate objects. If you want exactly six parts, brief a six-part chair in Step 2, the way the video's prompt does with its side frames, seat, backrest, and two rails. Extra parts are not a problem either. Step 6 shows how the web page groups them.

Split a Meshy model, and expect untextured parts. Auto Split works on a Meshy 6 or Meshy 7 model, textured or not, so the agent can run it on the chair from Step 2. The texture is dropped: each part comes back untextured, in one flat color. It takes a Meshy task as input, not a file you upload. Low-poly and Smart Topology outputs are not supported. On the Meshy website, the Auto Split button in the Print workspace also takes a plain-language list of the parts you want separated — the same describe-the-parts control this chat uses.

Check the result before moving on. Are the two side frames mirror images? Did the front rail come out as its own piece? If a cut landed somewhere else, tighten the part names or the construction wording in your Step 2 brief and re-split. Each run is a 10-credit job. If you only need the chair in pieces for 3D printing rather than for a web page, the Auto Split tutorial covers the print-focused path.

Step 5: Scale Up to a 20-Piece Collection

One chair is a demo. A catalog needs a collection that looks like it came from one designer. In the video the request is a single sentence:

@Meshy Now expand this into a collection of 20 furniture pieces across different categories. Show me the plan as a table first, then generate them.

Ask for the plan explicitly, as the last sentence does. The agent then writes a table with Category, Piece, and Design columns before it spends anything. Seating gets a dining chair, a lounge chair, a counter stool, a bench, a two-seat settee. Tables get a dining table, coffee table, side table, console. Each row carries the same design language as the hero chair: natural oak frames, curved walnut surfaces, rounded edges, slender proportions, matte finish. Read that table and edit it. Cutting a piece here is free. Cutting it after generation is not.

Once you approve the table, the agent generates the furniture 3D models one at a time. Each piece is one textured model: about two minutes and 30 credits, the same price whether the agent goes straight from text or makes a reference image first. When we ran a three-piece test through the plugin, the total came to exactly 90 credits.

When the batch finishes, the agent compares each result with its own plan and tells you what came out differently. In our test it reported that the side table had four legs instead of the three in the plan, and that the sideboard had handles the plan did not ask for. Treat that report as your review list: regenerate the pieces it flags, keep the rest.

Two things keep a collection looking like one designer made it. First, write the shared material rule into the plan itself, for example "unified by natural oak frames and curved walnut surfaces," so every piece inherits it instead of each prompt describing wood from scratch. Second, ask for predictable file names, like seating-dining-chair.glb. The web page you build in Step 6 reads these files by name.

Step 6: Build the Interactive 3D Product Page

Now the payoff, and the step where GPT-6 Astra does the most work. You have 20 GLB files, three color variants of the hero chair, and the chair in six named parts. Meshy made all of those. Astra now writes the code that shows them. Ask it to build the site:

Build an interactive furniture catalog site from the models in ./assets. Show all 20 pieces in a grid with category tabs (Seating, Sofas, Tables, Storage). Hover a card to see the 3D model turn. Give the Frame Chair its own page: drag to rotate, scroll to zoom, a color switcher for the three retextured variants, and a play button that assembles the six parts step by step.

In the video, the result is a page titled "Furniture. A new dimension." with a filterable grid and "hover for 3D, drag to explore" hints. The Frame Chair page is labeled "Assembly Study: Six Parts, One Chair." A slider walks through six stages, Layout, Frame, Front rail, Rear rail, Close frame, Seat, and Backrest, with each part tagged A to F. Every piece on the page is a Meshy model loaded in the browser.

We ran the same request on our three-piece test set and got the same kind of result without touching any code. The agent wrote a small site, started a local server, and handed back a link to open. The home page had category tabs and a card per piece. The chair page had a "Choose a finish" panel that swapped the three retextured files in place, a drag-to-rotate viewer, and a "Play assembly" button that flew the split parts in one at a time. Our test chair had split into 12 objects, and the agent labeled the page to match: "The split model contains 12 actual mesh parts. This is a visual assembly study." No credits were spent on this step. It is all code.

This is also how you build a 3D product configurator, the thing furniture and e-commerce sites pay agencies to make. A configurator is a product page where the shopper picks options and the 3D model updates. You already have the ingredients: one mesh, several texture variants that share it, and a viewer. The color switcher on the Frame Chair page is the simplest configurator there is. Add a second option set, say two leg finishes, and you have the real thing.

A few things to tell the agent so the page works well:

• Keep files small. Ask for models under a few megabytes each, or have the agent generate a lower-detail version for the grid and load the full model only on the product page.

• Use GLB for the web. It is the format browsers and 3D viewers expect. Ask for USDZ as well if you want iPhone users to place the chair in their room with AR.

• Name the parts in the file. The parts from Step 4 come out as separate objects. The agent can animate them individually only if they keep their names. If Auto Split gave you more pieces than you asked for, tell the agent which objects belong together ("the two stretchers are part of the frame") and it will group them into the six assembly steps.

Step 7: Review, Export, and Publish

Before anyone else sees the site, click through every product page yourself. Look for seams where the texture wraps, parts that do not quite meet after assembly, and cards that load slowly. Fix models with a retexture or a re-split. Fix loading with smaller files.

Then export what each channel needs. GLB for the site. USDZ for AR on Apple devices. If a piece is going to a factory or a print shop for a scale model, ask for STL or 3MF the way you would for any 3D printing job, and note that STL flattens the parts into one solid, so keep the multi-part version in GLB or 3MF. Meshy's export formats cover all of these.

Finally, get the site off your laptop. In our test the agent served it locally first, at a 127.0.0.1 address only you can open. Ask it where it can publish, or point it at a static host you already use and let it upload the folder. Once it is live, the catalog is a set of files you own. Add a piece by generating a new model and dropping it into the folder. Change a fabric by running one retexture. No 3D artist on retainer.

How to Write Prompts for 3D Furniture Modeling with AI

The armchair prompt in Step 2 is a template. Pull it apart and you get five slots that work for any piece of furniture:

1. Form. What it is and its silhouette: "a low, reclined lounge chair," "a round dining table."
2. Materials by part. Which wood or fabric goes where: "oak frame, walnut seat."
3. Construction. How the parts meet: "continuous side frames that become the legs," "two support rails beneath the seat."
4. Proportions. "Slender," "broad backrest," "low back."
5. Finish. "Matte," "oiled," "brushed."

Here is the formula applied to three more pieces from the video's own collection plan:

• Dining chair: "Slim oak frame with a gently curved walnut seat and back. Rounded edges, slender proportions, matte wood finish."

• Coffee table: "Low oval walnut top on curved oak legs. Rounded edges, matte finish."

• Console table: "Narrow walnut top with oak end frames and a lower shelf. Slender proportions, matte finish."

For AI furniture design across a whole set, the last two slots do the heavy lifting. Keep proportions and finish identical from piece to piece and the collection reads as one family, even when forms differ. Change them and you get a showroom of strangers. If you want to browse how other people describe furniture for Meshy, the furniture and home category has tens of thousands of finished models with their prompts.

How Much Does an AI 3D Product Catalog Cost in Credits?

Every task in this workflow has a published price, so you can budget a catalog before you start. These are the Meshy API rates at the time of writing, from the pricing reference:

Worked example for the hero chair: one textured model (30) plus three color variants (30) plus one named-parts split (10) comes to 70 credits. Twenty textured pieces with no variants come to about 600. A full collection with two fabric options each lands near 1,000. That is still a fraction of one commissioned model, but it adds up fast if you let the agent run without a plan, so approve the plan table in Step 5 before the batch starts.

Credits come from your Meshy subscription, and the plugin does not add a fee on top. The per-task breakdown is also in the credits cost guide. Plans and monthly credit amounts are on the Meshy pricing page.

FAQs

Can ChatGPT do 3D modeling?

Not on its own. ChatGPT cannot output a mesh file. With the Meshy plugin installed, it can run Meshy's 3D generation for you: you describe the object, the agent creates the task, and the finished model is saved to your computer. Install steps are in the ChatGPT and Codex guide linked earlier in this article.

Can ChatGPT design furniture?

Yes, in the sense that it can plan a collection, write detailed design prompts, and drive a 3D generator to produce each piece. In the video it turned one armchair prompt into a 20-piece plan and then generated the models through Meshy. The design judgment, which pieces to keep and which materials to pair, is still yours.

What is the difference between 3D product visualization and 3D product rendering?

Rendering makes still images or video from a 3D model. Visualization keeps the model interactive so a shopper can rotate, zoom, and switch options in the browser. Both start from the same textured model, so a workflow like this one feeds both.

Can I build a 3D product configurator with ChatGPT and Meshy?

Yes. A configurator needs one mesh, several texture variants that share it, and a web viewer. Meshy Retexture produces the variants without changing the geometry, and the agent can write the page with a color switcher, as it does for the Frame Chair in the video. Add more option sets the same way.

Do I need a Meshy API key to use the plugin in ChatGPT?

No. The plugin signs you in through your browser with a one-time code and stores the login on your computer. An API key is only needed if you call the Meshy API directly or use Meshy MCP in another agent.

How do I split a 3D model into named parts?

Ask the agent to run Meshy Auto Split in by_parts mode with the names, for example "left side frame, right side frame, seat, backrest, front rail, and rear rail." Say the mode explicitly, or the agent may run the default auto mode. Auto Split accepts up to ten names, costs 10 credits per run, and needs a Meshy 6 or Meshy 7 model made in Meshy. A textured model works, but the parts come back untextured. The names steer the cuts, so a model with more members than you listed can come back with more parts. The Auto Split API reference linked in Step 4 lists the modes and formats.

Do I need GPT-6 Astra to use the Meshy plugin?

No. The Meshy plugin works with any model you can pick in ChatGPT. GPT-6 Astra is the model the video uses and the one we tested with, because it is OpenAI's agent model, built for multi-step jobs like planning a collection and writing a web page. For a single model or a retexture, any model will do. For Steps 5 and 6, Astra is the better choice if your plan offers it.

Which 3D file format works best for a web product viewer?

GLB. It packs the mesh and textures into one file that browsers and 3D viewers load directly. Keep USDZ for AR on Apple devices and STL or 3MF for printing. Ask the agent for GLB by default and add the others only where you need them. If you would rather run the same jobs from a terminal, the Meshy CLI uses the same sign-in and credits.