| { |
| "as_of": "2026-10-24", |
| "strict_load": true, |
| "source_tensor_bytes_equal": true, |
| "smoke": { |
| "passed": true, |
| "device": "cpu", |
| "dtype": "original float32", |
| "encoder_layers": 12, |
| "decoder_layers": 28, |
| "rgb_fixture": "Deterministic synthetic RGB gradient/checkerboard, 224x224 input and 256x256 decoder output.", |
| "paper_quality_benchmark": false |
| }, |
| "checks": [ |
| { |
| "name": "all_selected_ema_payload_sha256", |
| "passed": true, |
| "tensors": 687 |
| }, |
| { |
| "name": "strict_complete_native_load", |
| "passed": true, |
| "tensors": 687, |
| "parameters": 503690496 |
| }, |
| { |
| "name": "full_shared_dino_encoder_depth", |
| "passed": true, |
| "layers": 12 |
| }, |
| { |
| "name": "full_decoder_depth", |
| "passed": true, |
| "layers": 28 |
| }, |
| { |
| "name": "normalized_generation_grid", |
| "passed": true, |
| "rms": 1.0641162395477295 |
| }, |
| { |
| "name": "semantic_and_reconstruction_features", |
| "passed": true |
| }, |
| { |
| "name": "native_feature_stats_and_grid_layout", |
| "passed": true, |
| "max_abs": 0.0, |
| "atol": 0, |
| "rtol": 0 |
| }, |
| { |
| "name": "full_encoder_independent_attention", |
| "passed": true, |
| "max_abs": 5.233287811279297e-05, |
| "atol": 0.0005, |
| "rtol": 0.0005 |
| }, |
| { |
| "name": "encoder_attention_reference_exercised", |
| "passed": true, |
| "calls": 12 |
| }, |
| { |
| "name": "full_depth_rgb_reconstruction", |
| "passed": true, |
| "shape": [ |
| 1, |
| 3, |
| 256, |
| 256 |
| ], |
| "unclamped_min": -0.24801495671272278, |
| "unclamped_max": 1.244450330734253 |
| }, |
| { |
| "name": "decoder_denormalize_unpatchify_rgb_contract", |
| "passed": true, |
| "max_abs": 0.0, |
| "atol": 0, |
| "rtol": 0 |
| }, |
| { |
| "name": "native_decode_noise_seed_reproducibility", |
| "passed": true, |
| "max_abs": 0.0, |
| "atol": 0, |
| "rtol": 0 |
| }, |
| { |
| "name": "original_decode_noise_changes_reconstruction", |
| "passed": true, |
| "mean_change": 0.03568131849169731 |
| }, |
| { |
| "name": "normalization_statistics_shape", |
| "passed": true |
| } |
| ], |
| "elapsed_seconds": 155.2128713130951, |
| "limitations": [ |
| "Synthetic RGB fixture, no reconstruction quality dataset or paper benchmark.", |
| "PR-DINOv2 tokenizer is paired with the separately released class DiT, not a complete Qwen3-8B multimodal language model.", |
| "Default decode_noise=True follows the release config; pass False for deterministic reconstruction." |
| ] |
| } |
