yuanze1024/ShaderAgent-MatScorer-Data

ShaderAgent MatScorer Data

Training and evaluation data for MatScorer, the material scorer used by ShaderAgent to rank candidate procedural materials against a region of a reference image. The trained adapter is yuanze1024/ShaderAgent-MatScorer-Qwen3VLReranker-LoRA.

The images are not uniformly licensed — read NOTICE before use. They are renders derived from BlenderKit, MatSynth and Tencent Hunyuan 3D, and the dataset as a whole is license: other. Only the annotations, viewer.html and ShaderAgent's own renders are CC BY 4.0. If you hold rights to an image here and want it removed, open an issue or a discussion.

What's in it

Each row is one listwise group: one query region, one positive candidate, and N negatives. Every candidate is three renders of the same material — plane, bright_ball, dark_ball, in that order.

config	label	splits	rows
synthetic	automatic, by retrieval similarity	train / val	11,573 / 1,318
human_pairs	human forced choice	train	364
eval	human preference ranking	tuning / method / stage2 / stage2_extra	161 / 644 / 205 / 43

synthetic is training stage S1 and human_pairs is S2. The eval config is held-out human judgment and was not trained on.

The eval splits differ in what the label compares:

• tuning — one method at different parameter-search iterations (17 groups).
• method — different methods on the same reference: ours_run1/2/3, blenderalchemy, vlmaterial, gemini, search (43 groups).
• stage2 / stage2_extra — forced choice over rollouts of one procedural material (74 + 17 groups). stage2_extra was mixed into the S2 training file upstream; its groups do not overlap stage2.

14,308 rows in total, over 59,441 distinct images (42.35 GiB, all PNG).

Browse it first

Open viewer.html: it renders a sample of every config with the query bbox drawn over the reference and the candidate views beside it. Clone the repo (or download viewer.html plus viewer_assets/) and open the file in a browser — straight off disk, no server and no network needed.

The Hub's Dataset Viewer shows image fields as strings, because the images are plain files rather than embedded in parquet. Embedding them would write each image once per reference — 268,191 references over 59,441 images — and inflate 42 GiB to roughly 205 GiB.

Record format

{
  "id": "synthetic/7d63b9b10b56a3d2c0f8dbaf",
  "query": {
    "image": "images/queries/synthetic/d6/d6dc5fadf3ec9b2f1d57025e08b046d0.png",
    "bbox": [30, 240, 392, 338],
    "image_size": [512, 512],
    "scene": "feminine-cosmetic-perfume",
    "camera": "Camera",
    "mask": "images/queries/masks/…png",
    "mask_coverage": 0.112213
  },
  "positive": {
    "source": "blenderkit",
    "material_id": "55acd530-cb85-58ae-bf54-c4bb2e779f71",
    "views": ["…/plane.png", "…/bright_ball.png", "…/dark_ball.png"],
    "retrieval_score": 0.443783
  },
  "negatives": [
    {"source": "matsynth", "material_id": "…", "views": ["…"], "difficulty": "hard", "retrieval_score": 0.916744}
  ],
  "label_source": "retrieval"
}

• bbox is pixel [x1, y1, x2, y2]. Resolutions are not uniform — 22 sizes occur, 512×512 covering 13,570 of 14,308 rows, the rest from 400×400 to 2048×2048, some non-square. Always read image_size alongside the bbox.
• label_source: retrieval, human_forced_choice, or human_ranking.
• source on a candidate: blenderkit, matsynth, procedural, or own_render (a method output in eval).
• difficulty on a synthetic negative: hard / medium / easy, from the retrieval neighbourhood it was drawn from.
• Negatives are order-independent. Training used InfoNCE over [positive, neg_1 … neg_N], so list order carries no meaning.
• eval rows also carry group.group_id, group.group_size, group.pair_count, group.pair_weight, group.kind (tuning, method, stage2, stage2_extra), and per-side uid / rank (rank 1 = preferred). Weight pairs by pair_weight to keep each group's total contribution equal.

Loading

from datasets import load_dataset
from huggingface_hub import snapshot_download
from PIL import Image
import os

ds = load_dataset("yuanze1024/ShaderAgent-MatScorer-Data", "synthetic")

root = snapshot_download("yuanze1024/ShaderAgent-MatScorer-Data", repo_type="dataset")
row = ds["train"][0]
ref = Image.open(os.path.join(root, row["query"]["image"]))
views = [Image.open(os.path.join(root, v)) for v in row["positive"]["views"]]

Image paths are repo-relative, so resolve them against a local snapshot. To fetch less than the full 42 GiB, pass allow_patterns. Everything human_pairs and eval reference is 2.1 GiB:

snapshot_download(
    "yuanze1024/ShaderAgent-MatScorer-Data", repo_type="dataset",
    allow_patterns=["annotations/*", "images/eval/**", "images/human_pairs/**"],
)

Both patterns are needed: eval/stage2 and eval/stage2_extra score rollouts of procedural materials, so their candidate views live under images/human_pairs/. The synthetic config is the other 40 GiB (images/cands/**, images/queries/**).

Layout

path	what	files
annotations/*.jsonl	the 14,308 records	7
images/cands/blenderkit/**	6,090 BlenderKit materials × 3 views	18,270
images/cands/matsynth/**	2,375 MatSynth materials × 3 views	7,125
images/cands/procedural/**	997 procedural materials × 3 views	2,991
images/queries/synthetic/**	reference renders	12,891
images/queries/masks/**	part masks for those references	11,543
images/human_pairs/**	S2 rollout renders, per material	5,585
images/eval/**	evaluation references and method outputs	1,036
viewer_assets/**	WebP thumbnails + cards.js for viewer.html	—

Query paths are sharded two levels deep by content hash, keeping any one directory well under the Hub's 10k-file guidance.

How it was built

Synthetic (S1). An object from HY3D-Bench is placed in a scene and one part's material is replaced with a known material. That render is the query; the bbox and mask come from the part's projection. Candidates are retrieved from the material catalogue: the material actually applied is the positive, and others are negatives bucketed into hard / medium / easy by retrieval similarity. These labels are automatic, so they carry retrieval bias — that is what S2 corrects.

Human pairs (S2). ShaderAgent's parameter search produces several rollouts of one procedural material against a reference, and a human picks which matches better. These labels teach the scorer what the retrieval score gets wrong.

Eval. Several methods (and several runs of ours) are run on the same reference, ranked by human preference, then expanded into pairs. Because the candidates are full method outputs rather than catalogue materials, this split measures what the scorer does at inference time.

Limitations

• Synthetic labels come from retrieval similarity, not human judgment, so a "negative" there may be a perfectly reasonable match.
• The human-labeled portion is small (364 training pairs, ~1,050 eval pairs) and reflects one annotator's preferences.
• Positives in synthetic are dominated by BlenderKit materials; MatSynth appears only as negatives.
• Reference images skew toward product-shot and interior scenes at 512×512.
• Baseline outputs were produced by the dataset author's own runs of those methods, not the original authors' reported results.

License

The annotations, viewer.html and ShaderAgent's own renders are CC BY 4.0. Everything else inherits its upstream's terms, and NOTICE states them per directory:

path	terms
annotations/, viewer.html, images/cands/procedural/**	CC BY 4.0
images/human_pairs/**/positive_*.png, neg_*.png	CC BY 4.0
images/eval/** except the two rows below	CC BY 4.0
images/cands/blenderkit/**, images/human_pairs/**/ref.png	renders of BlenderKit assets (Royalty Free or CC0 upstream)
images/cands/matsynth/**	renders of MatSynth (CC0 / CC-BY)
images/queries/**	Hunyuan 3D 2.1 Community License, and 90% also carry a BlenderKit material
images/eval/method/*/search/**	depicts a retrieved, usually BlenderKit, material
images/eval/refs/realworld/**	third-party photographs; no license granted

No upstream asset file is redistributed — no .blend files, meshes, texture maps or material archives, only PNG renders.

Citation

Please cite the ShaderAgent paper. If you use the MatSynth-derived renders, cite MatSynth as given in NOTICE.

@inproceedings{yuan2026shaderagent,
  title     = {ShaderAgent: Self-Evolving Agentic Procedural Material Generation},
  author    = {Yuan, Ze and Chen, Chia-hao and Zhang, Yuqing and Cao, Yan-Pei and Liang, Ding and Qi, Xiaojuan},
  booktitle = {SIGGRAPH Asia 2026 Conference Papers},
  year      = {2026},
  publisher = {ACM},
  doi       = {10.1145/3829340.3842164}
}