University of California San Diego
ECCV 2026
166 procedural materials generated from text prompts — hover any material to rotate it. Every one is an editable Blender node graph.
Artists do not learn from finished artworks alone. They learn by observing how other artists work: the decisions they make, the techniques they choose, and the reasoning behind those choices.
Material Apprentice is built on the same observation. It formulates procedural material generation as retrieval-time process reasoning: rather than synthesizing shader graphs directly, it retrieves expert demonstrations and reasons over the process artists used — the construction steps, parameter choices, and design intent behind them.
Expert workflows are represented as process traces extracted automatically from Blender tutorial videos. Given a request, a Process Synthesizer composes a trace aligned with the user's intent, and a Compiler grounds it into an executable Blender node graph — so every result is an editable procedural material, not a baked texture.
In an expert study with five professional Blender artists, materials generated by reflecting process expertise required fewer edits and matched professional design strategies more closely than graph-only synthesis. A user study with 150 participants further shows superior generation and editing performance over prior procedural systems.
The same process-centric reasoning supports text-to-material generation, image-conditioned synthesis, and text-based editing.
View this reel on Instagram
Artist Michael David Mayo reflects on the importance of learning from other artists.
Material Apprentice reasons over process traces derived from expert demonstrations — such as tutorial videos — capturing procedural sequencing, physical intuition, and the compositional design strategies artists use while building materials.
Technical overview
Blender procedural materials generated by Material Apprentice from the text prompt shown under each material. Hover to rotate.
We recruited five Blender artists (avg. 7.5 years of procedural-material experience) and evaluated 60 synthesized process traces from 30 prompts under two retrieval conditions: our process-trace retrieval versus a material-graph retrieval baseline. Artists first constructed a material graph by following each trace, then refined it for up to ten minutes to better match the prompt, while thinking aloud. We recorded every atomic edit, and an LLM analyzed the narrations with a fixed rubric to assign pairwise preferences and editing-friction labels.
| Metric | Process | Graph |
|---|---|---|
| Node edits ↓ | 2.42 | 3.42 |
| Connection edits ↓ | 5.42 | 7.42 |
| Parameter tweaks ↓ | 5.75 | 4.83 |
| Structural edits ↓ | 7.84 | 10.84 |
| Total edits ↓ | 13.58 | 15.67 |
| Criterion | Preference |
|---|---|
| Implementation clarity ↑ | 100% |
| Outcome quality ↑ | 92% |
| Pedagogical value ↑ | 84% |
| Procedural strategy ↑ | 83% |
| Parameter control ↑ | 83% |
| Final render ↑ | 83% |
| Signal | Δ (Process − Graph) |
|---|---|
| Confusion ↓ | −0.15 |
| Hesitation ↓ | −0.10 |
| Satisfaction ↑ | +0.40 |
| Completion ↑ | +0.25 |
Table 1: Expert study results. (Left) Average refinement edits per material — process-derived graphs need fewer structural and total edits; the slight increase in parameter tweaks is a favorable shift from repairing structure to tuning exposed controls. (Middle) Preferences from artists' think-aloud commentary; final render preference collected after refinement. (Right) Sentiment deltas after mapping low/medium/high to 1/2/3 — process-derived graphs reduce editing friction. In fully automatic generation over 50 prompts, three experts preferred process-conditioned outputs in 72% of paired comparisons.
Generate your own Blender procedural material without installing any code.
Step 1: Register your email
Material Apprentice is currently available as a research preview. To prevent abuse and manage compute resources, users must first register their email address before submitting jobs.
Academic and research users with .edu email addresses are preferred and are typically approved automatically.
Step 2: Run a generation
Once approved, you can submit a material prompt below. Material Apprentice will run the generation pipeline and email you a download link when the Blender asset is ready.
Material Apprentice uses your OpenAI API key only to execute the generation workflow associated with your request. We do not store, log, or reuse submitted API keys after the job completes.
Typical OpenAI API cost is approximately $2 USD per generation, although actual cost may vary depending on prompt complexity, reference image usage, retries, and model usage.
* Material Apprentice is provided as a research preview. OpenAI API usage is billed directly through your OpenAI account. While typical generations cost approximately $2 USD, actual costs may vary depending on model usage, request complexity, retries, and upstream API pricing. By using the demo, you acknowledge that you are responsible for any API charges incurred through your OpenAI account.
@inproceedings{gupta2026materialapprentice,
title = {Reflecting Process Expertise in Procedural Material Generation},
author = {Gupta, Kunal and Joshi, Gaurav and Chen, Yen-Ru and
Jain, Seemandhar and Mehta, Ishit and Chandraker, Manmohan},
booktitle = {European Conference on Computer Vision (ECCV)},
year = {2026}
}