Writing Prompts for AI 3D Generators

Where the advice comes from

Four vendors publish prompt guidance that we read, and each applies to its own tool. Advice for one model may not transfer to another, so check the source that matches the model you are using. Tencent’s page blocked our final check on 2026-11-08, so what follows about it is a paraphrase of an earlier read, not a quotation.

What they agree on

Across those sources, four pieces of advice repeat. Treat each as a claim by the vendor, not as a result.

• One object per prompt. Meshy, 3D AI Studio and Sloyd all warn against scenes or multiple objects. Tencent’s guide centers on one subject.
• Name real materials. Meshy: “Name real, physical materials.” 3D AI Studio lists material naming first under what improves a prompt. Sloyd lists “material hints”.
• State a style. All four include it.
• Prefer concrete words to adjectives of praise. Meshy’s rule against empty adjectives such as amazing and epic, and 3D AI Studio’s warning that “a cool thing” gives the AI nothing to work with, make the same point from different ends.

Meshy’s help article says what it expects to fail: smoke, glitter and “electric magic energy”, ultra-fine details such as individual strands of hair, and abstract concepts. Sloyd’s docs list its own weak spots: reflective or transparent materials, very thin details such as wires and chains, and faces and characters that “may look rough or ‘off’”. Those limitations are more useful than the tips, because they tell you when a prompt rewrite will not help and you should change route.

Where the advice disagrees

Two of these disagreements have practical consequences. The word-limit figures differ even inside one vendor, 3D AI Studio, so none of them is a specification. And a word in a prompt is not a control. Meshy’s API exposes target_polycount and pose_mode as real parameters, so where a setting exists, set it rather than hoping “5000 faces” or “T-pose” is obeyed, though the docs caution that the result can differ from the target. That is our editorial view, based on what the API documents; the vendors do not compare the two approaches.

On style references: naming a studio or franchise as a style is a separate question from whether the phrase works, and a generator allowing it does not settle what you can do with the output. See using AI 3D models commercially.

Starting templates by destination

These combine the vendors’ published patterns. They are starting points to test, not recipes known to work. The 3D prompt builder turns the same patterns into a fill-in form.

A prompt word cannot make a mesh game-ready or printable. Those claims are decided at the next step, which is why rigging and print checks still apply whatever the prompt said.

What Meshy’s published artist story shows

Meshy’s blog post (posted July 24, 2025) is a customer story by an artist, not documentation. She defines style pillars first (color, materials, silhouette, mood), then uses one fixed prompt template: an object “made from” named materials, then a style tag, proportions, a ban on modern plastics, a finish, and symmetry. Her tips are to keep the material language consistent, repeat core descriptors such as “brushed gold” across prompts, and not stack too many ideas. For image-to-3D she writes a shorter prompt to produce a reference image first. The post makes qualitative claims only. It gives no sample sizes, success rates or controlled comparison, and it promotes Meshy’s product.

Meshy’s July 2026 tutorial labels its 15 sample prompts “tested” but publishes no test data, and 3D AI Studio’s page headlines “real examples that work” with no outputs beside them, so neither supports a claim of effectiveness.

What each attempt costs

Prompt iteration spends credits, which is why test cheaply. Documented costs, accessed 2026-11-08:

• Meshy web app: Text to 3D is 20 credits on Meshy 6 and 10 on Meshy 6 Lite (the docs say Meshy 7.1 is not available in Text to 3D there). Failed generations are not charged and canceled tasks are refunded. Whether Lite output is good enough for judging a prompt is not documented. Meshy’s help center now says Text to 3D “is no longer available in the workspace” and happens in Meshy Agent, so check how your plan bills that route.
• Meshy API: Text to 3D Preview is 20 credits on meshy-7.1 (latest) and meshy-6, and 5 on meshy-6-lite, meshy-t2 and meshy-t2.5; Refine and Retexture are 10 credits at 2K/4K and 15 at 8K. Meshy says Preview and Refine are separate requests billed separately. The API limits prompt to 800 characters.
• Tripo API: H3.1 Text to 3D is listed as 10 credits without texture or 20 with standard texture, where 1 credit is $0.01 (developers.tripo3d.ai).
• 3D AI Studio: its docs FAQ gives Image or Text to 3D as 20 to 60 credits depending on the model (35 for Prism 3.1, 40 for Hunyuan 3.1 Pro), yet the same FAQ lists Tripo P2 at 70, the homepage FAQ says 25 to 35 and the pricing page says 20. Its docs say the cost is shown before you generate, so go by that figure.
• Sloyd Guest: one text-to-3D generation a day on the free tier.

See cost per usable asset for how prompt retries enter the cost, and Meshy pricing and credits for plan arithmetic.

How to test a prompt change yourself

Treat each advice item as a hypothesis. Pick one subject, write a baseline prompt, and change one thing at a time: length, an added material, an exclusion, a style tag. Keep the tool, plan, model version and settings fixed, record credits spent, and judge each result against your destination’s checks rather than a screenshot. Single generations vary, so one pair of results is not evidence; repeat before you conclude anything. The standard prompt set has the fields we plan to record.

A labeled budget example: one baseline plus three single changes, with three generations of each, is 12 generations. On published November 2026 rates that is 60 credits on Meshy API meshy-6-lite previews (12 × 5), 240 on meshy-6 previews (12 × 20), and 120 to 240 on Tripo’s API (12 × 10 to 20, or $1.20 to $2.40, on H3.1 without texture or with standard texture). Those figures cover preview or base generation only; texturing and retries are extra.

Decide before you generate what would count as better. For a game prop, that might be silhouette, a triangle count near budget, and no floating pieces. For a print, closed geometry and the thinnest feature at print size. Recording these first stops you from picking the prettiest render.

When to skip text

Several sources suggest an image reference helps for specific designs. Meshy’s help center recommends Image to 3D “for higher-level control over visual details”, and 3D AI Studio’s docs suggest generating and editing a reference image first. If you need a particular silhouette, see preparing images for image-to-3D. For choosing a tool, the text-to-3D shortlist compares them on documented features.

See also the AI 3D generation guides hub.