Add full RAI metadata (rai:dataSocialImpact rename + hasSyntheticData + prov:wasGeneratedBy) for NeurIPS 2026 E&D Track validator

croissant.json

@@ -7,6 +7,7 @@
     "conformsTo": "dct:conformsTo",
     "cr": "http://mlcommons.org/croissant/",
     "rai": "http://mlcommons.org/croissant/RAI/",
+    "prov": "http://www.w3.org/ns/prov#",
     "data": {
       "@id": "cr:data",
       "@type": "@json"
@@ -85,9 +86,15 @@
   "rai:dataBiases": "The 3D-FRONT source dataset consists of professionally designed Chinese residential interior layouts. Results may not generalize to: (1) non-Chinese architectural styles or culturally distinct interiors, (2) informal or low-income housing, (3) non-residential spaces (offices, hospitals, retail), (4) outdoor or mixed indoor-outdoor environments. All six generation models are Canny-conditioned; other conditioning types (depth, segmentation, sketch) are not evaluated. A single random seed (42) is used, limiting stochastic diversity analysis.",
   "rai:dataLimitations": "Stable Diffusion 1.5 generates at 512x512 resolution while the other five models generate at 1024x1024; cross-resolution comparison is noted as a limitation. Single seed (42) for all models. Monocular depth estimation is inherently scale-ambiguous and may introduce systematic errors. Synthetic ground truth may not capture all real-world material and lighting complexity. Only Canny edge conditioning is evaluated. Semantic masks are derived from render passes and may have minor aliasing artifacts at object boundaries.",
   "rai:dataUseCases": "Primary: benchmarking geometric fidelity of conditional image generation models for indoor scenes. Secondary: evaluating monocular depth estimation robustness across rendered vs. generated image domains; analysing ControlNet conditioning mechanism differences (channel concatenation vs. residual injection); architectural visualization quality assessment; 3D-aware image generation research; indoor scene understanding. Discouraged: do not use for generating photorealistic deceptive architectural renderings without explicit disclosure (the paper Section 6 notes that IGF-Bench could in principle facilitate such misuse, as with all generative-image benchmarks); do not use as the sole quality signal for production-grade 3D reconstruction without combining with held-out real-world data; do not use to evaluate generation outside the indoor-scene domain (the benchmark is restricted to enclosed indoor rooms).",
-  "rai:socialImpact": "This benchmark advances evaluation methodology for generative models in architecture and interior design. Potential positive impacts include improved quality control for AI-generated architectural imagery and more reliable synthetic data for downstream 3D reconstruction. Potential negative impacts are limited because the dataset contains only synthetic indoor scenes with no personal or sensitive content. Users should be aware that high-fidelity generated images could be misused for deceptive architectural renderings; the geometric metrics introduced here can serve as detection tools for such misuse.",
+  "rai:dataSocialImpact": "This benchmark advances evaluation methodology for generative models in architecture and interior design. Potential positive impacts include improved quality control for AI-generated architectural imagery and more reliable synthetic data for downstream 3D reconstruction. Potential negative impacts are limited because the dataset contains only synthetic indoor scenes with no personal or sensitive content. Users should be aware that high-fidelity generated images could be misused for deceptive architectural renderings; the geometric metrics introduced here can serve as detection tools for such misuse.",
   "rai:annotationsPerItem": "Each of the 3,600 ground-truth views directory contains 9 files: (1) RGB render (rgb_textured.png, 1024x1024 8-bit), (2) metric depth map (depth.exr, OpenEXR 32-bit float, in metres; canonical source), (3) 8-bit depth visualisation (depth.png, QA only), (4) NumPy float32 cache of depth.exr (depth_gt.npy, ~4 MB; for fast loading without an EXR decoder during LoRA training), (5) Canny edge map (canny.png, thresholds 100/200), (6) semantic segmentation mask with class IDs (semantic_id.png, 0=background, 1=floor, 2=ceiling, 3=wall, 4=opening, 5=structural element, 6=furniture), (7) colour-mapped semantic visualisation (semantic_mask.png, QA only), (8) 3D wireframe overlay (wireframe_3d.png, QA only), (9) camera parameters (camera.json: position, target, FOV=90, bounding box). Each of the 21,600 generated images has a corresponding DepthPro depth estimate (NPY) under depth_results/gen/<model>/depthpro/. Ablation subsets include Depth Anything V2 and ZoeDepth estimates.",
   "rai:dataReleaseMaintenancePlan": "Maintained for at least 5 years on HuggingFace (igfbench-neurips2026/IGF-Bench) and on the anonymous code mirror at anonymous.4open.science/r/IGF-Bench-Code (camera-ready: GitHub). Bug reports and questions are received via the HuggingFace dataset card discussion tab and via GitHub Issues on the code repository. Versioning follows semantic versioning conventions: minor metadata or documentation fixes ship as 1.x patch releases, schema-breaking changes (e.g. new room types, additional generation models, additional metrics) bump the major version, and older versions remain accessible on HuggingFace as immutable revisions. No deprecation is planned within the 5-year maintenance window.",
+  "rai:hasSyntheticData": true,
+  "prov:wasGeneratedBy": {
+    "@type": "prov:Activity",
+    "name": "IGF-Bench synthetic rendering and generation pipeline",
+    "description": "Ground-truth views were rendered from the 3D-FRONT scene corpus (Fu et al., 2021) at 1024x1024 using Blender 5.0 Cycles with physically based materials (4 calibrated viewpoints per room, FOV 90 degrees, eye-height 1.43 m). Canny edges were extracted from each GT render with thresholds 100/200 and used as ControlNet conditioning for six diffusion-based generation pipelines (Stable Diffusion 1.5, SDXL, Stable Diffusion 3.5, Flux.1 Dev, Hunyuan-DiT, Kolors) under uniform settings (seed 42, ControlNet strength 1.0, empty negative prompt). Monocular depth estimates were produced by Apple DepthPro (primary), with Depth Anything V2 and ZoeDepth as ablation alternatives. The pipeline is fully automated with deterministic seeds; no crowd workers or human annotators were involved. See DATASHEET.md Section 3 (Collection Process) for the complete provenance trace and the released code repository for runnable pipeline scripts."
+  },
   "distribution": [
     {
       "@type": "cr:FileObject",
@@ -637,5 +644,352 @@
       "encodingFormat": "application/json",
       "contentUrl": "evaluation/mde_ablation_results.json"
     }
+  ],
+  "recordSet": [
+    {
+      "@type": "cr:RecordSet",
+      "@id": "calibrated-views",
+      "name": "calibrated-views",
+      "description": "3,600 calibrated ground-truth views: 300 rooms x 3 complexity levels x 4 viewpoints. Each record contains all per-view assets.",
+      "field": [
+        {
+          "@type": "cr:Field",
+          "@id": "calibrated-views/room_id",
+          "name": "room_id",
+          "description": "3D-FRONT room identifier (UUID_RoomType-SceneID format)",
+          "dataType": "sc:Text",
+          "source": {
+            "fileSet": {
+              "@id": "gt-rgb-files"
+            },
+            "extract": {
+              "fileProperty": "fullpath"
+            },
+            "transform": {
+              "regex": "renders_textured/([^/]+)/L[012]_[a-z]+/view_[0-3]/rgb_textured\\.png"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "calibrated-views/level",
+          "name": "level",
+          "description": "Furniture complexity level: L0_empty (bare room), L1_basic (key furniture only), L2_full (fully furnished)",
+          "dataType": "sc:Text",
+          "source": {
+            "fileSet": {
+              "@id": "gt-rgb-files"
+            },
+            "extract": {
+              "fileProperty": "fullpath"
+            },
+            "transform": {
+              "regex": "renders_textured/[^/]+/(L[012]_[a-z]+)/view_[0-3]/rgb_textured\\.png"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "calibrated-views/view",
+          "name": "view",
+          "description": "Camera viewpoint index (view_0 through view_3)",
+          "dataType": "sc:Text",
+          "source": {
+            "fileSet": {
+              "@id": "gt-rgb-files"
+            },
+            "extract": {
+              "fileProperty": "fullpath"
+            },
+            "transform": {
+              "regex": "renders_textured/[^/]+/L[012]_[a-z]+/(view_[0-3])/rgb_textured\\.png"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "calibrated-views/rgb",
+          "name": "rgb",
+          "description": "Ground-truth RGB render (PNG, 1024x1024)",
+          "dataType": "sc:ImageObject",
+          "source": {
+            "fileSet": {
+              "@id": "gt-rgb-files"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "calibrated-views/depth_exr",
+          "name": "depth_exr",
+          "description": "Ground-truth metric depth map (EXR, 32-bit float, values in metres)",
+          "dataType": "sc:ImageObject",
+          "source": {
+            "fileSet": {
+              "@id": "gt-depth-files"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "calibrated-views/canny",
+          "name": "canny",
+          "description": "Canny edge map (PNG; low=100, high=200) used as ControlNet conditioning input",
+          "dataType": "sc:ImageObject",
+          "source": {
+            "fileSet": {
+              "@id": "gt-canny-files"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "calibrated-views/semantic_id",
+          "name": "semantic_id",
+          "description": "Semantic class label map (PNG; pixel values: 0=background, 1=floor, 2=ceiling, 3=wall, 4=opening, 5=structural element, 6=furniture). Primary planar classes 1/2/3 are used for L_plane and L_ortho.",
+          "dataType": "sc:ImageObject",
+          "source": {
+            "fileSet": {
+              "@id": "gt-semantic-id-files"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "calibrated-views/semantic_mask",
+          "name": "semantic_mask",
+          "description": "Semantic colour mask visualisation (PNG)",
+          "dataType": "sc:ImageObject",
+          "source": {
+            "fileSet": {
+              "@id": "gt-semantic-mask-files"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "calibrated-views/wireframe_3d",
+          "name": "wireframe_3d",
+          "description": "3D wireframe overlay render (PNG)",
+          "dataType": "sc:ImageObject",
+          "source": {
+            "fileSet": {
+              "@id": "gt-wireframe-files"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "calibrated-views/camera",
+          "name": "camera",
+          "description": "Camera parameters (JSON object with position, target, fov=90, and bounding box)",
+          "dataType": "sc:Text",
+          "source": {
+            "fileSet": {
+              "@id": "gt-camera-files"
+            }
+          }
+        }
+      ]
+    },
+    {
+      "@type": "cr:RecordSet",
+      "@id": "generated-images",
+      "name": "generated-images",
+      "description": "21,600 generated images produced by six Canny-conditioned diffusion models (6 models x 3,600 views).",
+      "field": [
+        {
+          "@type": "cr:Field",
+          "@id": "generated-images/room_id",
+          "name": "room_id",
+          "description": "3D-FRONT room identifier, matching calibrated-views/room_id",
+          "dataType": "sc:Text",
+          "references": {
+            "field": {
+              "@id": "calibrated-views/room_id"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "generated-images/level",
+          "name": "level",
+          "description": "Furniture complexity level (L0_empty, L1_basic, L2_full)",
+          "dataType": "sc:Text",
+          "references": {
+            "field": {
+              "@id": "calibrated-views/level"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "generated-images/view",
+          "name": "view",
+          "description": "Camera viewpoint index (view_0 through view_3)",
+          "dataType": "sc:Text",
+          "references": {
+            "field": {
+              "@id": "calibrated-views/view"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "generated-images/model",
+          "name": "model",
+          "description": "Generation model identifier: sd15 (Stable Diffusion 1.5, 512x512), sdxl (SDXL, 1024x1024), sd35 (Stable Diffusion 3.5, 1024x1024), flux1 (Flux.1 Dev, 1024x1024), hunyuan (Hunyuan-DiT, 1024x1024), kolors (Kolors, 1024x1024)",
+          "dataType": "sc:Text"
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "generated-images/image",
+          "name": "image",
+          "description": "Canny-conditioned generated image (PNG; 512x512 for sd15, 1024x1024 for all others; seed=42, cn_scale=1.0, empty negative prompt)",
+          "dataType": "sc:ImageObject",
+          "source": {
+            "fileSet": {
+              "@id": "gen-sd15-files"
+            }
+          }
+        }
+      ]
+    },
+    {
+      "@type": "cr:RecordSet",
+      "@id": "depth-estimates",
+      "name": "depth-estimates",
+      "description": "Monocular depth estimates for ground-truth renders and generated images. Primary estimator: DepthPro (25,200 files). Ablation estimators: Depth Anything V2 and ZoeDepth (subsets).",
+      "field": [
+        {
+          "@type": "cr:Field",
+          "@id": "depth-estimates/room_id",
+          "name": "room_id",
+          "description": "3D-FRONT room identifier",
+          "dataType": "sc:Text",
+          "references": {
+            "field": {
+              "@id": "calibrated-views/room_id"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "depth-estimates/level",
+          "name": "level",
+          "description": "Furniture complexity level",
+          "dataType": "sc:Text"
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "depth-estimates/view",
+          "name": "view",
+          "description": "Camera viewpoint index",
+          "dataType": "sc:Text"
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "depth-estimates/image_source",
+          "name": "image_source",
+          "description": "Whether depth was estimated from a ground-truth render (gt) or a generated image (gen)",
+          "dataType": "sc:Text"
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "depth-estimates/gen_model",
+          "name": "gen_model",
+          "description": "Generation model name (sd15, sdxl, sd35, flux1, hunyuan, kolors); only populated when image_source=gen",
+          "dataType": "sc:Text"
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "depth-estimates/mde_method",
+          "name": "mde_method",
+          "description": "Monocular depth estimation method: depthpro (primary), dav2 (Depth Anything V2, ablation), zoedepth (ZoeDepth, ablation)",
+          "dataType": "sc:Text"
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "depth-estimates/depth_npy",
+          "name": "depth_npy",
+          "description": "Estimated depth map stored as a NumPy array (.npy, float32)",
+          "dataType": "sc:DataDownload",
+          "source": {
+            "fileSet": {
+              "@id": "depth-gen-depthpro-files"
+            }
+          }
+        }
+      ]
+    },
+    {
+      "@type": "cr:RecordSet",
+      "@id": "evaluation-metrics",
+      "name": "evaluation-metrics",
+      "description": "Pre-computed evaluation metrics per model per complexity level. Four geometric fidelity metrics (lower is better) and two perceptual metrics.",
+      "field": [
+        {
+          "@type": "cr:Field",
+          "@id": "evaluation-metrics/model",
+          "name": "model",
+          "description": "Generation model identifier",
+          "dataType": "sc:Text",
+          "references": {
+            "field": {
+              "@id": "generated-images/model"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "evaluation-metrics/level",
+          "name": "level",
+          "description": "Furniture complexity level",
+          "dataType": "sc:Text"
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "evaluation-metrics/L_plane",
+          "name": "L_plane",
+          "description": "Planarity error: mean deviation of fitted planes on structural surfaces (wall, floor, ceiling) from ground truth. Lower is better.",
+          "dataType": "sc:Float"
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "evaluation-metrics/L_ortho",
+          "name": "L_ortho",
+          "description": "Orthogonality error: deviation of wall-floor and wall-ceiling angles from 90 degrees. Lower is better.",
+          "dataType": "sc:Float"
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "evaluation-metrics/L_edge",
+          "name": "L_edge",
+          "description": "Edge straightness error: deviation of structural edges from straight lines. Lower is better.",
+          "dataType": "sc:Float"
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "evaluation-metrics/L_vp",
+          "name": "L_vp",
+          "description": "Vanishing point consistency error: angular deviation of detected vanishing points from ground-truth camera geometry. Lower is better.",
+          "dataType": "sc:Float"
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "evaluation-metrics/FID",
+          "name": "FID",
+          "description": "Frechet Inception Distance (perceptual, lower is better)",
+          "dataType": "sc:Float"
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "evaluation-metrics/LPIPS",
+          "name": "LPIPS",
+          "description": "Learned Perceptual Image Patch Similarity (perceptual, lower is better)",
+          "dataType": "sc:Float"
+        }
+      ]
+    }
   ]
 }