Fix viewer file:// loading; correct image licensing in NOTICE

NOTICE

@@ -48,14 +48,37 @@ MatSynth is CC0 with a CC-BY subset. Cite the upstream work:
 
 
 images/cands/procedural/**
-images/human_pairs/**
 --------------------------
 Renders of procedural Blender materials, including node graphs that
-ShaderAgent itself wrote. These renders are released under CC BY 4.0.
+ShaderAgent itself wrote. Released under CC BY 4.0.
+
+
+images/human_pairs/**
+---------------------
+This directory mixes two things, and they do not carry the same terms.
+
+  images/human_pairs/<material>/ref.png            renders of BlenderKit
+                                                     materials; see the
+                                                     BlenderKit section above
+
+      455 files, one per material directory. These are the *reference*
+      images — the material each pair was asked to reproduce. Each was made
+      by appending a BlenderKit material datablock into a scene and
+      rendering it, so treat these the way you would treat anything under
+      `images/cands/blenderkit/`: a rendered depiction, not a redistribution
+      of the asset, and not CC BY 4.0. (All 455 directory names are exact
+      BlenderKit catalogue entries, and each render's own build log records
+      the material being appended from that catalogue's `.blend`.)
 
-Some of the material names under `images/human_pairs/` come from procedural
-material sources in the BlenderKit catalogue; as above, only renders are
-published.
+  images/human_pairs/<material>/positive_*.png     CC BY 4.0
+  images/human_pairs/<material>/neg_*.png          CC BY 4.0
+
+      5,130 files. These are ShaderAgent's own procedural attempts at
+      reproducing the reference — node graphs the system wrote itself.
+
+The `eval` config's `stage2` and `stage2_extra` splits score rollouts of
+procedural materials, so their candidate views also live in this directory and
+the same split applies.
 
 
 images/queries/**
@@ -63,6 +86,14 @@ images/queries/**
 Reference ("query") renders: an object from HY3D-Bench placed in a scene, with
 one part's material swapped, plus the corresponding part mask.
 
+Each of these renders therefore combines three third-party things: the object
+mesh, the background scene, and **the material applied to the swapped part**.
+That material is a BlenderKit material in 11,558 of the 12,891 query renders
+(90%) and a procedural material in the remaining 1,333 — it is the same
+material as the record's `positive` candidate, whose `source` field names the
+catalogue. So the BlenderKit terms described above apply to most of these
+renders as well, on top of the object terms below.
+
 The 3D objects come from Tencent Hunyuan 3D. Their use is governed by the
 **Tencent Hunyuan 3D 2.1 Community License**, which, among other terms:
 
@@ -78,7 +109,8 @@ https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1
 The meshes themselves are **not** redistributed here — only renders.
 
 The scene files used as backgrounds are third-party and are **not**
-redistributed. Only the renders are published.
+redistributed. Only the renders are published. The same applies to the 52
+background scenes behind `images/human_pairs/**/ref.png`.
 
 
 images/eval/**
@@ -86,7 +118,17 @@ images/eval/**
 Renders produced by ShaderAgent and by the baseline methods it was compared
 against (BlenderAlchemy, VLMaterial, a Gemini-generation variant, and a
 retrieval-only "search" variant), all run by the dataset author on the author's
-own hardware. Released under CC BY 4.0.
+own hardware. Released under CC BY 4.0, with one exception:
+
+  images/eval/method/*/search/**                  depicts a retrieved material
+
+      114 files. The "search" variant does not synthesise a material; it
+      retrieves one from the material catalogue and renders it. The render is
+      the author's own (none of them is byte-identical to any catalogue
+      render), but the material depicted is whatever was retrieved, which is
+      usually a BlenderKit material. Treat these the way you would treat
+      `images/cands/blenderkit/`. The other methods generate or optimise their
+      own materials, so their renders are CC BY 4.0.
 
 `images/eval/refs/**` holds the reference images the evaluation was run
 against, in four groups: `blender` (renders), `sd` and `aigen` (generated
@@ -110,11 +152,16 @@ Summary table
                                                   (Royalty Free or CC0 upstream)
   images/cands/matsynth/**                      renders of MatSynth (CC0 / CC-BY)
   images/cands/procedural/**                    CC BY 4.0
-  images/human_pairs/**                         CC BY 4.0
+  images/human_pairs/**/ref.png                 renders of BlenderKit assets
+  images/human_pairs/**/positive_*, neg_*       CC BY 4.0
   images/queries/**                             renders of Hunyuan 3D objects
                                                   (Tencent Hunyuan 3D 2.1
                                                    Community License applies)
-  images/eval/**                                CC BY 4.0
+                                                  carrying a BlenderKit material
+                                                   in 90% of cases
+  images/eval/**                                CC BY 4.0, except:
+  images/eval/method/*/search/**                  depicts a retrieved
+                                                   (usually BlenderKit) material
   images/eval/refs/realworld/**                 third-party photographs,
                                                   no license granted
 

README.md

@@ -47,21 +47,37 @@ procedural materials against a region of a reference image. The trained adapter
 lives at
 [yuanze1024/ShaderAgent-MatScorer-Qwen3VLReranker-LoRA](https://huggingface.co/yuanze1024/ShaderAgent-MatScorer-Qwen3VLReranker-LoRA).
 
-**Read [NOTICE](NOTICE) before use.** The images are renders derived from
-several upstream collections with different terms (BlenderKit, MatSynth,
-Tencent Hunyuan 3D), which is why the dataset is labeled `license: other`.
-The reference photographs under `images/eval/refs/realworld/` were collected
-from assorted sources and carry no license grant here. If you hold rights to
-any image in this dataset and want it removed, open an issue or a discussion
-and it will be removed promptly.
+**Read [NOTICE](NOTICE) before use — the images are not uniformly licensed.**
+They are renders derived from several upstream collections with different terms
+(BlenderKit, MatSynth, Tencent Hunyuan 3D), which is why the dataset is labeled
+`license: other`. In particular:
+
+- **BlenderKit terms reach further than the `images/cands/blenderkit/`
+  directory.** At least 18,839 files (32%) are renders of BlenderKit materials,
+  counting the candidate views, the 455 `human_pairs/*/ref.png` references and
+  the 114 `eval/method/*/search/` renders. On top of that, 11,558 of the 12,891
+  query renders (90%) show a BlenderKit material applied to the object. No
+  BlenderKit asset file is redistributed — only renders — but do not treat these
+  as openly licensed.
+- **`images/eval/refs/realworld/`** holds photographs collected from assorted
+  sources; no copyright is claimed and no license is granted for them.
+- **`images/queries/`** renders contain Tencent Hunyuan 3D objects, whose
+  community license excludes the EU, the UK and South Korea from its territory.
+
+What is cleanly CC BY 4.0: the annotations, `viewer.html`, and the renders that
+ShaderAgent produced itself — `images/cands/procedural/`, the 5,130
+`human_pairs` rollouts, and the non-`search` method outputs under `images/eval/`.
+
+If you hold rights to any image here and want it removed, open an issue or a
+discussion and it will be removed promptly.
 
 ## Browse it first
 
 Open **[`viewer.html`](viewer.html)** — a self-contained page that renders a
 sample of every config with the query bbox drawn over the reference image and
 the candidate views beside it. Clone the repo (or download `viewer.html` plus
-`viewer_assets/`) and open the file in a browser; it needs no server and no
-network.
+`viewer_assets/`) and open the file in a browser — straight off disk, no server
+and no network needed.
 
 The Hub's own Dataset Viewer shows the `images/**` fields as *strings*, because
 the images are stored as plain files rather than embedded in parquet. Embedding
@@ -128,9 +144,11 @@ the evaluation data rather than with `human_pairs`. Their groups do not overlap
 ```
 
 - **`bbox` is pixel `[x1, y1, x2, y2]`** on the query image. Reference
-  resolutions are not uniform (512×512 dominates, but 400×400 up to 2048×2048
-  occur), so **always read `image_size` alongside it** rather than assuming a
-  size. All 14,308 boxes are within bounds.
+  resolutions are not uniform: 22 distinct sizes occur, 512×512 covering 13,570
+  of the 14,308 rows, the rest ranging from 400×400 to 2048×2048 and including
+  non-square ones (1800×800, 639×1512, 1920×1256). **Always read `image_size`
+  alongside the bbox** rather than assuming a size or a square aspect. All
+  14,308 boxes are within their image bounds.
 - **`label_source`** is `retrieval` (automatic), `human_forced_choice`, or
   `human_ranking`.
 - **`source`** on a candidate is `blenderkit`, `matsynth`, `procedural`, or
@@ -170,7 +188,9 @@ ref = Image.open(os.path.join(root, row["query"]["image"]))
 views = [Image.open(os.path.join(root, v)) for v in row["positive"]["views"]]
 ```
 
-To fetch only part of the 42 GiB, use `allow_patterns`:
+To fetch only part of the 42 GiB, use `allow_patterns`. Everything the
+`human_pairs` and `eval` configs reference is **2.1 GiB** — all 6,621 of their
+images plus the annotations:
 
 ```python
 snapshot_download(
@@ -179,6 +199,11 @@ snapshot_download(
 )
 ```
 
+Note that `eval/stage2` and `eval/stage2_extra` score rollouts of procedural
+materials, so their candidate views live under `images/human_pairs/` rather than
+`images/eval/` — both patterns are needed. The `synthetic` config is the other
+40 GiB (`images/cands/**` and `images/queries/**`).
+
 ## Contents
 
 | path | what | files |
@@ -191,7 +216,7 @@ snapshot_download(
 | `images/queries/masks/**` | part masks for those references | 11,543 |
 | `images/human_pairs/**` | S2 rollout renders, per material (train + eval) | 5,585 |
 | `images/eval/**` | evaluation references and method outputs | 1,036 |
-| `viewer_assets/**` | WebP thumbnails + `cards.json` for `viewer.html` | — |
+| `viewer_assets/**` | WebP thumbnails + `cards.js` for `viewer.html` | — |
 
 59,441 images, 42.35 GiB, all PNG. Query paths are sharded two levels deep by
 content hash to keep any one directory well under the Hub's 10k-file guidance.

viewer.html

@@ -54,20 +54,21 @@ footer{color:var(--muted);font-size:11px;text-align:center;padding:24px}
   <div class="controls" id="controls">
     <span class="sub">config</span>
     <button data-kind="cfg" data-v="all" class="on">all</button>
-    <span class="sub" style="margin-left:10px">label</span>
+    <span class="sub" id="lab-label" style="margin-left:10px">label</span>
     <button data-kind="lab" data-v="all" class="on">all</button>
     <input id="q" type="search" placeholder="filter by id / scene / material / group…">
   </div>
 </header>
 <main>
   <div class="stats" id="stats">loading…</div>
-  <div id="cards"><div class="skeleton">loading viewer_assets/cards.json …</div></div>
+  <div id="cards"><div class="skeleton">loading viewer_assets/cards.js …</div></div>
 </main>
 <footer>
   images shown are lossy WebP thumbnails · full-resolution PNGs are the
   <code>images/…</code> paths referenced by the annotations ·
   <a href="https://huggingface.co/datasets/yuanze1024/ShaderAgent-MatScorer-Data">dataset card</a>
 </footer>
+<script src="viewer_assets/cards.js"></script>
 <script>
 const $ = (s, r=document) => r.querySelector(s);
 const $$ = (s, r=document) => [...r.querySelectorAll(s)];
@@ -126,7 +127,8 @@ function buildFilters(){
   const cfg = [...new Set(CARDS.map(c => CFG_OF_SPLIT(c.split)))].sort();
   const lab = [...new Set(CARDS.map(c => c.label_source))].sort();
   const host = $('#controls');
-  cfg.forEach(v => host.insertBefore(mkBtn('cfg', v), host.querySelector('span[style]')));
+  const anchor = $('#lab-label');
+  cfg.forEach(v => host.insertBefore(mkBtn('cfg', v), anchor));
   lab.forEach(v => host.appendChild(mkBtn('lab', v)));
   $$('button').forEach(b => b.onclick = () => {
     $$(`button[data-kind=${b.dataset.kind}]`).forEach(x => x.classList.toggle('on', x === b));
@@ -140,17 +142,20 @@ function mkBtn(kind, v){
   return b;
 }
 
-fetch('viewer_assets/cards.json').then(r => {
-  if (!r.ok) throw new Error('cards.json missing — was the dataset packaged?');
-  return r.json();
-}).then(d => {
-  CARDS = d.cards || [];
+// cards.js assigns this global. A <script src> works from file://, which is how
+// the dataset card tells people to open this page; fetch() would be blocked.
+try {
+  const d = window.__MATSCORER_CARDS__;
+  if (!d || !Array.isArray(d.cards)) {
+    throw new Error('viewer_assets/cards.js did not load — keep it next to this file');
+  }
+  CARDS = d.cards;
   buildFilters();
   render();
-}).catch(e => {
-  $('#cards').innerHTML = `<div class="skeleton">could not load cards.json: ${esc(e.message)}</div>`;
+} catch (e) {
+  $('#cards').innerHTML = `<div class="skeleton">could not load the card data: ${esc(e.message)}</div>`;
   $('#stats').textContent = '';
-});
+}
 $('#q').addEventListener('input', ev => { TEXT = ev.target.value.trim().toLowerCase(); render(); });
 </script>
 </body>