PromptModel

This document covers everything from first-run activation through the troubleshooting table. If you only have five minutes, read Quick Start and Writing Prompts.

⤓ Installation

Import once. The Editor window lights up immediately.

From the Unity Asset Store

Dependencies

PromptModel pins three Unity Package Manager dependencies. They are added automatically on first import and visible in Window → Package Manager → In Project:

What gets installed

Assets/PromptModel/
  Editor/                 // All Editor-side C#: window, services, importers, helpers
    UI/                   // Window, banner, drop zone, presenter
    Services/             // HTTP client, auth store, generation coordinator
    Importers/            // GLB → Prefab pipeline orchestrator
    Helpers/              // MaterialApplier, LodGenerator, PrefabWrapper
    DTOs/                 // Request/response types
  Documentation/          // You are here
  Samples/                // Showcase scene + sample prefabs
  Generated/              // Output folder; created on first generation
  EULA.md                 // License terms
  README.md               // 30-second intro
Project

⚡ Quick Start

Zero to your first generated prefab in under three minutes (plus generation time).

⚙ How It Works

From prompt to prefab in one round-trip.

Text pipeline

Image pipeline

1. You click Generate. The window posts { invoice, text_prompt OR image_url, quality_preset, seed } to https://model.promptmotion.dev/prod/generate with your invoice as authentication context.
2. The backend returns 202 Accepted with a request_id and starts the generation in the background.
3. The window polls GET /prod/status/{request_id} every two seconds, updating the progress bar.
4. For text-mode, the server runs your prompt through FLUX schnell (a fast text-to-image model) to get a reference image, then through TRELLIS (Microsoft's image-to-3D model). Image-mode skips the FLUX step.
5. When the GLB is ready, the window downloads it via a HMAC-signed URL into Assets/PromptModel/Generated/_raw/.
6. Our import pipeline applies materials, generates LOD0/1/2/3, centres the pivot, and saves a Unity Prefab next to the raw GLB.
7. If Spawn into scene is checked, the prefab is instantiated in your active scene at world origin and selected.

⚡ Powered By

Open-source models, third-party hosting, full attribution.

Microsoft TRELLIS

The 3D-generation model is TRELLIS — published by Microsoft Research at CVPR 2025, released under the MIT License. TRELLIS produces a textured mesh from a single reference image and is the strongest publicly-available image-to-3D model as of mid-2026.

FLUX schnell

For text-mode generations, an intermediate text-to-image step uses Black Forest Labs' FLUX schnell. The image it produces is consumed by TRELLIS and not surfaced separately to your project.

fal.ai

Both models are hosted by fal.ai, a serverless AI inference provider. PromptModel's backend is a thin wrapper around fal's API plus our own auth, quota, and caching layers. fal's Terms of Service apply to your use of PromptModel; the relevant pass-through clauses are summarised in EULA § 3.

Whinarn UnityMeshSimplifier

LOD0/1/2/3 generation uses UnityMeshSimplifier (MIT License) at 100% / 50% / 25% / 10% triangle counts respectively.

Unity gltFast

The downloaded .glb is imported by Unity's official gltFast package. From there our import pipeline takes over for material extraction, LOD generation, and prefab wrapping.

✏ Writing Prompts

Clear, single-subject, concrete. AI 3D models are a different prompt regime than text-to-image.

What works well

• Single concrete subject — a prop, an object, a creature. Not "a battle scene" or "the inside of a tavern".
• Material details — "iron-banded", "carved wood", "rusty", "polished marble".
• Style hints — "stylized low-poly", "PBR realistic", "cartoon".
• Game-ready scope — barrels, weapons, tools, furniture, plants, simple creatures.

What doesn't work as well

• Multiple subjects in one prompt — "a sword and a shield" usually produces a fused mess. Generate them separately.
• Hero characters with detailed faces — TRELLIS handles props much better than people. Try image-mode with a clean reference instead.
• Photorealistic cinematic detail — the output is meant for 3D viewports, not VFX shots. Set the Quality dropdown to High Detail if you need more triangles.
• Branded or copyrighted characters — this is also against the EULA.

Examples

Quality preset

🎨 Image Input

Drop an image; TRELLIS reconstructs a 3D model from it. Frequently produces sharper outputs than text-mode for visually-distinct subjects.

Three input paths

1. Texture2D from Project view — drag a Texture2D asset into the drop zone. Requires Read/Write Enabled in the texture's import settings (a clear inline error is shown if not, and the asset is pinged).
2. OS file from Explorer / Finder — drag a .png / .jpg / .jpeg / .webp file straight from your operating system.
3. Click to browse — click anywhere on the empty drop zone to open a native file picker.

Limits

What makes a good reference image

• Single subject filling most of the frame — TRELLIS treats the image as one object. If there are two, expect a fused result.
• Clean, neutral background — white, grey, or transparent works best. Busy backgrounds confuse silhouette extraction.
• Even, soft lighting — harsh shadows get baked into the mesh's albedo and look wrong from other angles.
• Three-quarters view — pure front or pure profile gives TRELLIS less to extrapolate from. A 30–45° offset is the sweet spot.
• Resolution sweet spot 512–1024 px — below that loses detail; above isn't beneficial for current TRELLIS quality.

📁 Generated Files

Where everything lands and what each file is for.

Assets/PromptModel/Generated/
  _raw/
    req_xxxxxxxx.glb                  // Raw GLB downloaded from the backend
  wooden_barrel/                      // Folder named after your prompt / Name field
    wooden_barrel.prefab              // The prefab to drop into your scene
    wooden_barrel.metadata.json       // Sidecar: prompt + seed (powers "Generate similar")
    wooden_barrel 1.prefab            // Second generation of the same name (Unity's space-suffix)
    wooden_barrel 1.metadata.json
    materials/
      wooden_barrel_mat_00.mat        // Native URP Lit / Standard material
      wooden_barrel_mat_00_basemap.png
      wooden_barrel_mat_00_metallicglossmap.png
    lods/
      geometry_0_LOD1.asset           // 50% triangles (LOD asset names track the source mesh)
      geometry_0_LOD2.asset           // 25% triangles
      geometry_0_LOD3.asset           // 10% triangles
  wooden_barrel_v2/                   // Variations from "Generate similar" get their own folder
    wooden_barrel_v2.prefab           // Original is conceptually v1; variations start at _v2
    wooden_barrel_v2.metadata.json
    ...
  knight_helmet/
    knight_helmet_untextured.prefab   // Prefab suffix when fallback grey material was applied
    knight_helmet_untextured.metadata.json
    ...
Project

Naming

The basename is derived in this order: explicit Name field → your prompt → image filename → pmod_{request_id}. The result is sanitised: lowercase, spaces → underscores, only [a-z 0-9 _ -] kept, max 64 characters.

If a prefab with the same basename already exists, Unity's AssetDatabase.GenerateUniqueAssetPath appends a space-prefixed counter ( 1, 2, …) so the new prefab lands as wooden_barrel 1.prefab in the same folder. We never overwrite a previous prefab. Variations from Generate similar follow a different convention — each gets its own {name}_v{N}/ folder starting at _v2.

The _untextured suffix

If TRELLIS produces a GLB with no PBR textures (rare but possible for very abstract prompts), the pipeline applies a neutral grey fallback material and suffixes the prefab name with _untextured. You'll also see a [PromptModel] WARN: GLB had no PBR textures in the Editor Console.

What's safe to delete

• Whole folder — deletes the prefab, materials, and LOD meshes for that generation. Safe.
• _raw/ — deletes the original GLBs but keeps the prefabs. Safe; you lose the ability to re-run the pipeline locally on those models, but the prefabs themselves still work.
• Individual .mat or .png — the prefab will go pink (missing material). Don't do this unless you're replacing the file.

💡 About PBR materials and lighting

Why generated models can look like they have shadows baked in — and how to compensate.

PromptModel uses Microsoft TRELLIS for 3D generation. Like all current AI 3D models, TRELLIS bakes lighting and shadows into the model's texture during generation. The albedo (base color) texture includes the implicit lighting from the reference image, rather than being a clean unlit surface color.

What this means for your scenes

• Generated models work best in scenes with soft, diffuse lighting — HDRI skyboxes, area lights, or carefully tuned ambient.
• Avoid harsh single directional lights — they double-up with the baked-in shadows on the texture.
• Adjust via the Material Inspector if needed: lower metallic, raise smoothness, or apply a slight tint to compensate.

For cinematic / hand-painted-quality PBR projects

• Unity's built-in Window → Rendering → Delighting tool can help separate baked light from albedo.
• Third-party tools like Marmoset Toolbag or Adobe Substance Sampler do this professionally.
• We're exploring automated delighting as a v1.1 feature.

➤ Auto-Spawn

When checked, the prefab drops into your active scene at world origin as soon as it's ready.

The Spawn into scene when done checkbox below the Generate button controls this. Default on; preference persists across Editor restarts.

What it does

1. Instantiates the prefab in the currently-active scene.
2. Positions it at world origin (0, 0, 0).
3. Selects the new GameObject so it shows up in the Hierarchy.
4. Frames it in the Scene view (SceneView.FrameSelected).

When it's skipped

🔒 Activation

Your Asset Store invoice is your license. No signup, no API key, no account.

Paste your Unity Asset Store invoice number into the field at the top of the PromptModel window and click Verify. Your invoice activates the asset across all your installations — there's no machine-specific binding.

Your purchase includes 50 generations every month. Cached prompts return instantly without using your allowance.

What gets sent to the server

• Your invoice number (HTTPS, JSON body).
• A client version string for analytics + compatibility checks.

Not sent: name, email, IP (other than what's inherent to TLS), Unity Editor analytics, scene contents, or any telemetry beyond the above. See Privacy.

Forgetting an invoice

Choose Window → PromptModel → Forget Invoice. This clears the locally stored invoice.

⏲ Monthly Allowance

Each successful generation uses one of your monthly generations. Cache hits and failures cost zero.

What counts as a generation

A "generation" is a single submission to POST /prod/generate that:

• Reaches a terminal state of completed;
• Produced a downloadable GLB; AND
• Was a fresh inference run, not a cache hit.

What does NOT count

• Failed generations — FLUX_FAILED, TRELLIS_FAILED, PROVIDER_RATE_LIMIT, etc.
• Generations cancelled by you before reaching completed.
• Cache hits (cached for 30 days; identical (prompt OR image_url, quality_preset, seed) within that window returns the previous prefab for free).
• Image uploads (rate-limited separately to 30/hour).
• Status polls.

How the monthly allowance works

Your purchase includes 50 generations every month. Your next allowance arrives at the start of the next calendar month — unused generations do not roll over.

The History panel

The bottom of the window shows your last 10 generations and the most recent quota line. The numbers refresh after every generation and every status poll — no need to manually refresh.

⚙ Generate Similar (Variations)

Run the same prompt 4 times in parallel, pick the best one.

TRELLIS is non-deterministic — the same prompt with different seeds produces meaningfully different geometry. The Generate similar button lets you fan out a single prompt into 4 parallel variations, preview each as a thumbnail in a modal, then import only the ones you like.

How to launch

Two entry points, both producing the same result:

• From the History panel: each row in the bottom panel has a Generate similar button. The button is disabled for entries created before this feature shipped (the original prompt wasn't recorded back then).
• From the Project view: right-click a PromptModel-generated prefab in the Project window → PromptModel: Generate similar. The menu is greyed out on prefabs that didn't come from PromptModel.

What the modal shows

A utility window with up to four slots in a 2×2 grid. Each slot fills in independently as its variation completes:

• Thumbnail — best-effort rendering from Unity's asset preview system. If the preview can't be generated for some reason, you'll see a checkmark placeholder — the variation is still importable.
• Status badge — Queued → Generating → Ready (or Failed).
• Seed label — the random seed used for that variation, in case you want to reproduce one later.
• Keep this variation checkbox — tick the ones you want to import.

After you click Import selected, each ticked variation runs through the same import pipeline as a regular generation. The new prefabs land in Assets/PromptModel/Generated/ with names like wooden_barrel_v2.prefab, wooden_barrel_v3.prefab — the original is conceptually v1, so variations start at _v2.

Cost & quota

Each variation uses one of your monthly generations, just like a regular generation. A 4-variation batch uses 4 generations. The modal shows a quota counter that updates after submission.

What you can keep

Only ticked + completed slots are imported. Failed slots can't be ticked. If you close the modal without ticking anything, no quota is wasted on import — the generations were already consumed when the batch was submitted, but no asset clutter is added to your project. Unticked variations are deleted from the local cache when you close the modal.

Variations of variations

Imported variations are themselves PromptModel-generated prefabs — so they get their own metadata sidecar and are valid sources for another Generate similar round. The modal carries forward the original prompt, not the variation's seed, so you keep generating off the same intent rather than drifting through transformed-of-transformed.

📄 Where's My Invoice?

Two places. Either is fine.

Option 1: Asset Store website

1. Sign in at assetstore.unity.com.
2. Top right → your avatar → My Account → My Orders.
3. Find PromptModel in the order list.
4. The invoice number looks like IN_xxxxxxxxxxxxxxx.

Option 2: order confirmation email

Search your inbox for "Unity Asset Store order confirmation". The invoice number is in the email body and on the attached PDF.

Unity's own help article is at support.unity.com — How can I get an invoice for an Asset Store purchase. The "Where do I find my invoice?" link in the activation banner points there.

💡 Troubleshooting

Every error code the window can show, what it means, and what to do.

The colour bar on each entry corresponds to the in-window action:

• Pink — you need to do something (re-enter invoice, replace image, edit prompt).
• Yellow — we'll auto-retry; just wait.
• Blue — informational; no action required.

Auth errors

Quota / rate-limit errors

Input / validation errors

Provider / pipeline errors

Service / network errors

🔒 Privacy

Short version: we don't sell your data, prompts and images aren't used to train models, and we keep almost nothing about you.

What we collect

• Your invoice ID (sent in the JSON body of authenticated requests).
• Your prompts (text strings).
• Your reference image bytes when you submit image-mode generations.
• A client version string.

What we don't collect

• Your name, email address (other than what's in your support email if you write us), or phone number.
• Your IP address (other than what's inherent to TLS termination on Cloudflare).
• Unity Editor analytics, scene contents, or telemetry beyond the above.
• Any cookies in the Editor (no browser involved).

Retention

• Generated GLBs and uploaded reference images: cached for 30 days. After that they're deleted from disk.
• Per-generation audit log entries (cost telemetry, error codes, durations): retained for 90 days for operational debugging and refund accounting.
• Invoice + binding records: retained for the lifetime of the service.

Third-party processing

Your prompts and reference images are forwarded to fal.ai and Microsoft TRELLIS for inference. fal.ai's Terms of Service govern their handling of your data; they retain inputs only as needed to perform inference. We do not authorise (and they do not perform) model training on customer inputs.

GDPR-style rights

The publisher is in Hungary (EU). Email ragendom@gmail.com for: a copy of all data we hold for your invoice, deletion, or correction. We respond within 30 days. See EULA § 9 for the full text.

⟳ Updating PromptModel

Standard Asset Store update flow. Your data persists.

1. Open Window → Package Manager → My Assets.
2. If a new version is available, the Update button appears next to PromptModel.
3. Click Update, then Import. Accept the import dialog.

What persists across updates

• Your invoice and binding state (in EditorPrefs).
• The History panel's last 10 generations.
• Your auto-spawn preference.
• Anything in Assets/PromptModel/Generated/ — package updates never touch your generated outputs.

What gets replaced

• Source code under Assets/PromptModel/Editor/.
• Documentation, EULA, README.
• Sample scene and bundled prefabs.

❓ FAQ

The questions support gets most often.

📬 Support

One inbox. We read everything.

What to include

The faster we can reproduce the issue, the faster we can fix it. Please include:

• Your invoice ID (IN_xxxxxxxxxxxxxxx).
• The diagnostic ID from any error message you've seen (looks like diag_xxxxxxxxxxxx).
• Your Unity Editor version and operating system.
• The prompt or image filename involved.
• A short description of what you were trying to do, and what happened instead.
• Console output (copy/paste the relevant lines — the window logs [PromptModel]-prefixed messages there).

What we can't help with directly

• Asset Store refunds — those go through Unity's standard process at Unity Support. We can confirm refund-related auth state, but we can't issue the refund ourselves.
• fal.ai or TRELLIS bugs — we relay them upstream but can't fix model behaviour ourselves. We can often restore the affected generation to your monthly allowance on your behalf.
• Custom integrations / SLAs — v1 doesn't offer enterprise tiers. Email anyway and we'll see if we can help.

Other ways to follow PromptModel

• Asset Store update notes — check the listing for new releases and known issues.
• The publisher's portfolio site (linked in the Asset Store listing).