Using Marso Studio
This page describes how to use Marso Studio.
At this stage, the web renderer we use isn't representative of the quality of the materials. We recommend evalutating in your own renderer.
Marso Studio is a node-based web application.
The workflow is straightforward:
Upload Node — Drag or select your .glb file. The viewer will display your mesh with its current base texture so you can confirm the asset loaded correctly.
You can also use one of our example assets to run your first generation, these can be found when you click upload model and head to the Example Assets tab.
Generate PBR Node — Connect to the upload node and run the prediction. The I2M model processes your base texture and produces the PBR texture pack.
Review & Download — Inspect the results in the 3D viewer and download the output textures (albedo, roughness, IOR, metallic).
You can toggle on Show Normal Map in the 3D viewer within Marso Studio. Please note: this does not affect the generated PBR output, only the render within the viewer.
What to Expect — Strengths & Limitations ✅
We want to be upfront about where the model excels and where it currently has rough edges.
Where it works well
Well-defined surfaces — objects with clear material properties (stone, wood, plastic, ceramic) tend to produce strong, convincing PBR.
Clean photogrammetry — if the scan is good, the PBR prediction will be good. Garbage in, garbage out applies here.
Known limitations
Metallic prediction can be unstable. Metallic is the hardest property to predict. The model can sometimes misclassify regions as metallic or non-metallic, particularly on ambiguous surfaces (e.g. painted metal, wet stone, dark plastics). This is an active area of improvement.
Coverage artefacts on complex geometry. We've tuned the number of inference steps to optimise for online performance — keeping prediction times reasonable for a web app. The trade-off is that complex geometry can sometimes show subtle artefacts, especially in high-detail or concave regions where coverage is harder. This is something we can resolve with a higher step count, which produces cleaner, more complete results. The high-step-count mode isn't currently exposed in the web app, but we've included some comparisons below so you can see the difference.
Feedback
This is a pre-release build and your feedback is incredibly valuable. If you run into issues, see unexpected results, or have suggestions — please let us know. The more specific you can be (which asset, what went wrong, screenshots if possible) the more it helps us improve.
Last updated
