schema_version: spatialgen.protocols.v1
benchmark: SpatialGen-Bench
benchmark_version: v1.1
sample_count: 470

runtime:
  framework: ProVisE
  repository: https://github.com/XuWang917/ProVisE
  revision: fdee9debe21e6b838d41f3ea9f2d6146bd1d24c1

policy:
  construct_before_model_evaluation: true
  same_protocol_across_models: true
  deterministic_parser_preferred: true
  original_metric_preserved: true

tasks:
  counting:
    capability: perception
    input_mode: single
    protocol: instance_marker_count
    response: one green star on each target instance
    parser: green_connected_components
    metric: exact_count

  depth:
    capability: perception
    input_mode: single
    protocol: relative_depth
    response: grayscale depth map with brighter pixels closer to the camera
    parser: grayscale_coordinate_sampling
    metric: accuracy

  orientation:
    capability: perception
    input_mode: single
    protocol: direction_grid
    response: fill the corresponding outer cell of a 3x3 direction grid
    parser: blue_cell_to_direction
    metric: accuracy

  size:
    capability: perception
    input_mode: single
    protocol: grounded_size_comparison
    response: mark the compared objects and show visible extent or measurement evidence
    parser: constrained_generated_image_readout
    metric: accuracy

  relationship:
    capability: understanding
    input_mode: single
    protocol: binary_color_presence
    response: subject/object color masks for true, grayscale scene for false
    parser: color_presence
    metric: accuracy

  perspective:
    capability: understanding
    input_mode: single
    protocol: direction_grid
    response: fill the 3x3 grid cell for the target direction from the reference viewpoint
    parser: blue_cell_to_choice
    metric: accuracy

  mental_modeling:
    capability: understanding
    input_mode: multiple
    protocol: top_down_scene_model
    response: top-down layout with camera positions and facing directions
    parser: constrained_generated_image_readout
    metric: accuracy

  grounding:
    capability: understanding
    input_mode: single
    protocol: point_marker
    response: one prominent cyan point inside the referred object
    parser: cyan_marker_centroid
    metric: point_in_mask

  multihop:
    capability: reasoning
    input_mode: single
    protocol: state_edit
    response: apply the requested spatial operations while preserving unrelated content
    parser: constrained_generated_image_readout
    metric: accuracy

  prediction:
    capability: reasoning
    input_mode: single
    protocol: future_trajectory
    response: draw one physically consistent trajectory to the predicted target
    parser: constrained_generated_image_readout
    metric: hole_distance_partial_credit

  feasibility:
    capability: reasoning
    input_mode: single
    protocol: fit_rendering
    response: depict the queried placement at realistic scale with visible clearance or collision
    parser: constrained_generated_image_readout
    metric: accuracy

  affordance:
    capability: interaction
    input_mode: single
    protocol: region_mask
    response: binary mask of the target affordance region
    parser: binary_mask
    metric: mask_precision

  navigation:
    capability: interaction
    input_mode: single_with_candidates
    protocol: state_matching
    response: generate the next view after the indicated camera movement
    parser: clip_candidate_matching
    metric: accuracy

  trajectory:
    capability: interaction
    input_mode: single
    protocol: trajectory
    response: one continuous red path from the start point to the target
    parser: red_path_extraction
    metric: dfd_success_at_0.4