{
  "__task_type_notes__": ["estimation_or_reconstruction: RGB is used as the condition to generate a non-RGB modality; this covers reconstruction-style tasks such as RGB-to-3D/spatial information.","conditional_control: a non-RGB modality is used as the condition to generate RGB or another non-RGB modality.","conditional_control+narrow_generation: caption and a non-RGB modality are used together to generate RGB; this is both conditional control and caption-driven narrow generation.","general_generation: caption alone is used as the condition to generate any modality; currently not planned for training because RGB can be generated first and other modalities can then be reconstructed/estimated.","narrow_generation: caption alone is used as the condition to generate RGB; base text-to-image/text-to-video models belong to this category.","inverse_rendering: RGB plus another modality are used together as conditions to generate a non-RGB modality.","sparse adapter note: adapter_task selects the concrete sparse training path used by scripts/launch_scripts/train_sparse_structure_models.sh.","sparse adapter note: condition_preprocess records derived features such as PARE crops, full RGB tensors, or normalized object point clouds.","sparse adapter note: validated default Heads store their deployment choice directly in recommended_checkpoint; optional experimental Heads omit that key.","audio reconstruction note: cropped visual source cues can be used as estimation_or_reconstruction conditions for waveform/audio targets such as 175-184.","audio sparse note: 175-184 is the real scene/source audio; audio localization uses [175-184, 0-2] as audio+full-RGB conditions to predict a source box, audio segmentation uses the same conditions to predict the source mask, while audio classification uses 175-184 alone. 185-194 caption/TTS audio is kept out of the default sparse training set."],
  "normalization": {
    "default_mode": "log",
    "modalities": {
      "0-2": "linear",
      "90-92": "linear",
      "122-124": "linear",
      "150": "linear"
    },
    "channel_ranges": {
      "0": [0.0,255.0],
      "1": [0.0,255.0],
      "2": [0.0,255.0],
      "3": [0.0,255.0],
      "32": [-3.599609375,1.8564453125],
      "33": [0.0,1.0],
      "34": [0.58447265625,67.8125],
      "35": [0.0,1.0],
      "36": [-1.2443599700927734,1.1317600011825562],
      "37": [-1.3442360162734985,1.3391599655151367],
      "38": [-1.4413199424743652,1.495519995689392],
      "39": [-20.703125,16.5],
      "40": [0.0,255.0],
      "41": [-20.234375,18.109375],
      "42": [-15.78125,16.03125],
      "64": [0.0,20.0],
      "65": [0.0,127.0],
      "66": [0.0,341.0],
      "67": [0.0,255.0],
      "68": [0.0,255.0],
      "69": [0.0,255.0],
      "90": [0.0,255.0],
      "91": [0.0,255.0],
      "92": [0.0,255.0],
      "95": [0.0,5.0],
      "96": [0.0,32.0],
      "97": [-0.830078125,0.193939208984375],
      "98": [0.1759033203125,1.0],
      "99": [0.0,255.0],
      "100": [0.0,224.0],
      "101": [0.0,231.0],
      "103": [0.0,5.0],
      "111": [-1.0,417.0],
      "122": [0.0,255.0],
      "123": [0.0,255.0],
      "124": [0.0,255.0],
      "126": [0.30760085582733154,100.0],
      "127": [0.42604973912239075,14.270201683044434],
      "128": [0.36309295892715454,132.59178161621094],
      "129": [0.07100828737020493,2.3783669471740723],
      "130": [0.13239231705665588,50.0],
      "131": [-1.0,19.0],
      "132": [18.075801849365234,100.0],
      "133": [0.6000000238418579,2.740000009536743],
      "134": [0.30000001192092896,0.6800000071525574],
      "135": [0.0,1.0],
      "136": [4.0,247.0],
      "143": [0.0,213.0],
      "144": [0.0,1.0],
      "145": [0.0,1.0],
      "146": [0.0,1.0],
      "147": [0.0,1.0],
      "148": [0.0,1.0],
      "149": [0.0,1.0],
      "150": [-1.0,1.0],
      "151": [0.0,0.0],
      "152": [0.0,0.0],
      "153": [0.0,0.0],
      "154": [1.0,1.0],
      "155": [0.0,1.0],
      "156": [0.0,1.0],
      "157": [0.0,1.0],
      "161": [0.0,0.90576171875],
      "162": [0.0,0.6943359375],
      "163": [0.0,0.64306640625],
      "164": [0.0,0.93310546875],
      "165": [0.0,0.7646484375],
      "166": [0.0,0.7607421875],
      "167": [0.153076171875,0.88623046875],
      "168": [0.0,1.0],
      "169": [0.0,1.0],
      "170": [0.0,0.39599609375],
      "171": [0.0,0.36865234375],
      "172": [0.0,0.35693359375],
      "173": [0.0,1.0],
      "195": [-1.0,26.0]
    }
  },
  "wsg_v1": {
    "Channel_Head_000000": {
      "task_type": "estimation_or_reconstruction","pair_exists": true,"adapter_task": "smpl","target_density": "sparse","condition": "0-2","condition_description": "0-2: RGB crop encoded as PARE features","condition_preprocess": "rgb_crop_pare","target": "72-87,90-93","target_description": "72-87 and 90-93: SMPL/SMPLer-X body pose, shape, and related sparse human-structure parameters used by the ScoreHMR adapter",
      "recommended_checkpoint": "train_log_sparse/Channel_Head_000000/20260810_184401_wsv_structured_scorehmr/scorehmr/checkpoints/run/model-step_0010000.pt"},
    "Channel_Head_000001": {
      "task_type": "estimation_or_reconstruction","pair_exists": true,"adapter_task": "bbox","target_density": "sparse","condition": "0-2","condition_description": "0-2: full RGB image resized for bbox prediction","condition_preprocess": "rgb_full_224","target": "88-89","target_description": "88-89: multi-person bounding boxes represented by the bbox adapter; boxes are derived from MediaPipe person/id visibility channels when building the dataset","dense_to_sparse_variants": [{"modality": "88-89_bbox_pseudocolor_mask","storage": "data/sparse_structure/flux2_bbox_pseudocolor_mask","encoding": "black-background filled instance rectangles with a fixed left-to-right pseudocolor palette, encoded by the FLUX.2 VAE","decode": "prototype-color segmentation followed by one bounding rectangle per recovered instance region"}],
      "recommended_checkpoint": "train_log_sparse/Channel_Head_000001/20260811_144000_bbox_pretrained_center_gaussian/ckpt/model-best.pt"},
    "Channel_Head_000002": {
      "task_type": "estimation_or_reconstruction","pair_exists": true,"adapter_task": "dwpose","target_density": "sparse","condition": "0-2","condition_description": "0-2: RGB crop encoded as PARE features","condition_preprocess": "rgb_crop_pare","target": "64-66","target_description": "64: DWPose person id | 65: DWPose keypoint id | 66: DWPose confidence/score",
      "recommended_checkpoint": "train_log_sparse/Channel_Head_000002/20260811_135035_dwpose_kinematic_rotations/ckpt/model-best.pt"},
    "Channel_Head_000003": {
      "task_type": "estimation_or_reconstruction","pair_exists": true,"adapter_task": "mediapipe","target_density": "sparse","condition": "0-2","condition_description": "0-2: RGB crop encoded as PARE features","condition_preprocess": "rgb_crop_pare","target": "95-98","target_description": "95: MediaPipe person id | 96: MediaPipe joint id | 97: relative z depth | 98: visibility confidence",
      "recommended_checkpoint": "train_log_sparse/Channel_Head_000003/20260811_135035_mediapipe_kinematic_rotations/ckpt/model-best.pt"},
    "Channel_Head_000004": {
      "task_type": "conditional_control","pair_exists": false,"adapter_task": "obj","target_density": "sparse","condition": "111,112,114-116","condition_description": "111/112/114-116: object id, class value, and vertex xyz fields converted into a normalized object/group point cloud","condition_preprocess": "object_pointcloud_norm","target": "112-121","target_description": "112-121: sparse object/animal structure payload; current adapter trains a BFS skeleton tree from class=3 joints and class=4 bones"},
    "Channel_Head_000005": {
      "task_type": "inverse_rendering","pair_exists": true,"adapter_task": "scene_audio_rgb_to_semantic_instance_localization","target_density": "sparse","condition": ["175-184","0-2"],"condition_description": "175-184: real scene/source video audio | 0-2: full RGB image used as the visual grounding condition","condition_preprocess": "scene_audio_feature+rgb_full","target": "195_box","target_description": "195-derived sound-source box supervision from pipeline29/audio mask; used for audio-source localization",
      "recommended_checkpoint": "train_log_sparse/Channel_Head_000005/20260811_183000_clip_frozen_mask_fused_final/ckpt/model-best.pt"},
    "Channel_Head_000006": {
      "task_type": "conditional_control","pair_exists": true,"adapter_task": "scene_audio_to_semantic_category","target_density": "sparse","condition": "175-184","condition_description": "175-184: real scene/source video audio used as the only condition","condition_preprocess": "scene_audio_feature","target": "195","target_description": "195: AVIS semantic category id supervision from pipeline29; used for audio category recognition/classification",
      "recommended_checkpoint": "train_log_sparse/Channel_Head_000006/20260811_175200_prototype_temporal5_video_balanced/ckpt/model-best.pt"},
    "Channel_Head_000007": {
      "task_type": "estimation_or_reconstruction","pair_exists": false,"adapter_task": "crop_rgb_to_scene_audio_waveform","target_density": "sparse","condition": "0-2","condition_description": "0-2: cropped RGB/source image used as the visual condition for real scene/source audio generation","condition_preprocess": "rgb_crop_audio_source","target": "175-184","target_description": "175-184: real scene/source video audio generated or reconstructed in waveform/audio feature space"},
    "Channel_Head_000008": {
      "task_type": "conditional_control+narrow_generation","pair_exists": true,"adapter_task": "scene_audio_to_semantic_mask_dense","target_density": "dense","condition": ["175-184","CurrentFrame_RGB_Condition"],"condition_description": "175-184: frame-level AST audio feature condition | CurrentFrame_RGB_Condition: current RGB latent used as a trainable spatial condition lane","condition_preprocess": "scene_audio_feature_dense+current_rgb_condition","target": "195","target_description": "195: dense semantic/source mask latent predicted from frame-level audio condition and current RGB condition; used for audio-source segmentation/inverse rendering"},
    "Channel_Head_000009": {
      "task_type": "estimation_or_reconstruction","pair_exists": true,"adapter_task": "rgb_audio_to_semantic_mask_dense_prototype_ce","target_density": "dense","condition": ["175-184","CurrentFrame_RGB_Condition"],"condition_description": "CurrentFrame_RGB_Condition: current RGB latent used as the primary spatial estimation condition | 175-184: frame-level AST audio feature injected as one global auxiliary condition token","condition_preprocess": "current_rgb_condition+scene_audio_feature_dense_global_token","target": "195","target_description": "195: dense semantic/source mask estimated directly from the current RGB frame with current-frame audio as auxiliary evidence; trained with the target-195 prototype CE recipe"},
    "Channel_Head_000010": {
      "task_type": "estimation_or_reconstruction","pair_exists": true,"adapter_task": "rgb_to_semantic_mask_audio_region_selector","target_density": "dense","condition": ["0-2","175-184"],"generation_condition": "0-2","categorical_refiner_condition": "175-184","condition_description": "0-2: RGB drives candidate channel-195 mask geometry | 175-184: frame audio drives the default categorical Head, which selects zero, one, or multiple sounding candidate regions and assigns their semantic colors","condition_preprocess": "rgb_dense_generation+audio_ast_candidate_set_selection","target": "195","target_description": "195: audio-selected semantic source mask; RGB generates candidate regions and the default audio-conditioned categorical Head performs multi-region selection and exact category coloring",
      "recommended_checkpoint": "train_log_sparse/Channel_Head_000010/20260823_205300_head010_rgb195_frozen_category_audio_selector/ckpt/final_step_0001000/pytorch_lora_weights.safetensors"}
  }
}
