AI Texture Generation for Robot Vision Training Data: A Practical Guide to Domain Randomization

What domain randomization is, and why texture is one of its highest-leverage parameters

Domain randomization is the standard technique for closing the "sim-to-real gap" — the well-documented problem where a vision model trained entirely in simulation underperforms once it meets real cameras, real lighting, and real object surfaces. The original NVIDIA domain-randomization paper describes the approach as randomizing simulator parameters such as lighting, pose, and object textures so the model learns essential task features rather than one specific rendering; NVIDIA's Isaac Sim Replicator documentation shows the same pattern in synthetic-data workflows.

Texture is not a minor parameter in that mix. In the original NVIDIA CVPR paper that extended domain randomization to real-object detection (Tremblay et al., 2018), removing randomized textures from the objects of interest measurably hurt detection accuracy in the paper's own ablation study — texture randomization sat alongside lighting as one of the most important levers in the whole pipeline.

This shows up constantly in more recent robotics research, too:

• RoboTwin 2.0 (2025), a bimanual manipulation benchmark, built its own texture library by generating 20,000 AI images from 1,000 surface descriptions collected through LLM prompting and web crawling, then human-filtering them down to roughly 11,000 usable background textures — purely to get enough material variety into training.
• SHOP-VRB2, a household-object manipulation benchmark, explicitly randomizes objects across plastic, metal, glass, rubber, and wood finishes — the exact "same table, different material" pattern that shows up in almost every domain-randomization paper.
• Industrial bin-picking research consistently flags textureless and reflective metallic parts as one of the hardest unsolved problems in robotic grasping vision, precisely because appearance shifts so dramatically with material and lighting.

In other words, if your object detector or grasp-pose estimator only ever sees one finish of one object, it hasn't actually learned the object. It's learned a photo.

The real bottleneck isn't the 3D model. It's the textures.

Most teams in this space already have what they need geometrically — a CAD export, a scanned mesh, or a downloaded asset. What they don't have is a fast way to put a dozen or a hundred plausible, production-quality material variants on that same mesh. The available options each have a real cost:

• Manual texturing in Substance Painter — hours of work per variant, and it requires an actual 3D artist on the team. This doesn't scale to "50 materials × 200 object classes."
• Scanned CC0 texture libraries (think ambientCG, Poly Haven, Quixel Megascans) — free and photoreal, but a fixed catalog. You get whatever's already been scanned, not the specific "oxidized green-patina copper" or "scratched matte powder-coated steel" variant your test matrix actually calls for, and you still have to fit and tile it to your model's UVs by hand.
• Procedural shader graphs — infinite variation, but they require node-graph/shader expertise that most robotics and ML engineers simply don't have time to learn.
• Building your own generative pipeline, as the RoboTwin 2.0 team did, with an LLM prompting Stable Diffusion and a human filtering the output. It works — but it took a research team real engineering time to stand up, and what it produced were flat color images, not full PBR material sets with usable normal, roughness, and metallic maps.

This is the gap a fast, prompt-driven, PBR-complete texture generator is built to close.

Step by step: from a 3D model to a domain-randomization-ready texture pool

Start with a UV-unwrapped 3D model. This is already a required step in standard simulation-asset pipelines — NVIDIA's SimReady UV layout guidance notes that simulation assets with complex texture map layouts require a UV mapping process before those maps can be accurately projected onto the surface. TextureFast assumes this same starting point, so it drops directly into an existing workflow rather than adding a new one.

1. 1

   Upload the model

   GLB, GLTF, OBJ, and FBX are all supported — GLB is the fastest path since it's a single self-contained binary file.

2. 2

   Describe the material, or hand it a reference image

   "Brushed stainless steel with glare details on the surface", "brand new brown leather, and gray metallic legs", "modern light brown wood with copper metallic parts" — or upload a photo and let prompt expansion fill in the rest.

3. 3

   Generate the full PBR set in one click

   TextureFast first produces a Base Color/Albedo texture from the description, then extracts the complete material — Albedo, Normal, Height, Roughness, Metallic, and AO — at up to 4K resolution.

4. 4

   Repeat with a new material description

   Because each Base Color texture generation takes roughly 40 seconds, producing 10, 30, or 100 material variants of the same mesh for a randomization pool is a matter of minutes, not a multi-day art-production task.

5. 5

   Feed the resulting texture set into your randomizer

   Drop the maps into Omniverse Replicator's randomization parameters, Unity Perception's TextureRandomizer/MaterialRandomizer asset lists, or any custom domain-randomization script.

Worked example: one table mesh, five materials, one afternoon

Take the canonical domain-randomization example: a single table mesh that needs to train a detector to recognize "table" regardless of finish.

Five prompts, five PBR-complete generations, each producing Albedo/Normal/Height/Roughness/Metallic/AO maps on the same UV-mapped mesh — done in well under fifteen minutes of active work. Scale that pattern across every object class in a bin-picking, warehouse, or household-object training set, and the texture-sourcing bottleneck that used to require either a 3D artist or a bespoke generative-AI pipeline shrinks to an afternoon task.

How this lines up with Isaac Sim, Omniverse SimReady, and Unity Perception

This isn't a coincidental fit — the output format matches the industry spec closely:

• NVIDIA's SimReady materials guidance recommends the PBR Roughness/Metallic workflow, texture maps baked at 4K or smaller, and 1–2K as a baseline for smaller or distant objects. That lines up with the map set and resolution ceiling TextureFast outputs.
• GLTF, OBJ, and FBX — plus GLB as the binary GLTF handoff TextureFast supports — line up with common USD import paths listed in NVIDIA's Omniverse format guidance.
• Unity Perception's TextureRandomizer is explicitly built to sample from a developer-supplied list of texture assets. It provides the randomization mechanism, not the textures — sourcing that list is left entirely to the developer, which is exactly the step this workflow accelerates.

In practice, that means a TextureFast-generated material set can go straight from download into a USD/MDL conversion step or directly into a Unity/Blender/Unreal-based synthetic-data pipeline without reformatting.

A privacy note that matters more in this industry than most

Unreleased robot hardware, unannounced consumer products, and proprietary industrial parts are everywhere in this field, and a lot of generative AI tooling is unclear (at best) about what happens to uploaded geometry. TextureFast's stated approach — no training on uploaded assets, no asset exposure, NDA-safe, fully private generation — is a genuinely meaningful differentiator for any team working with geometry that can't leak, not just a checkbox feature.

Comparison: how the options actually stack up

Pricing in the context of a real texture-randomization workload

A small research team running domain randomization across, say, 15–20 object classes with 10–20 material variants each is generating somewhere in the 150–400 textured-model range — comfortably inside TextureFast's Starter-to-Pro tier volumes. A team building a larger SimReady asset catalog across hundreds of SKUs is closer to the Business tier's ~2,461 textured models/month. Either way, the per-asset economics compare favorably against either an artist's hourly rate or the engineering time of standing up a custom generative pipeline.