SnoopyFan commited on
Commit
e69966d
·
verified ·
1 Parent(s): 8c8b94e

Add WorldCompass Sana training artifacts

Browse files
rebuttal_data/worldcompass_sana/wc_sana_smoke_f242310_20260801e/worldcompass_provenance.json ADDED
@@ -0,0 +1,105 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "adaptation_name": "WorldCompass objective adapted to official 4-step SANA-WM streaming",
3
+ "candidate_gradient_accumulation": 2,
4
+ "candidates_per_group": 6,
5
+ "chunk_layout": [
6
+ 0,
7
+ 4,
8
+ 7,
9
+ 10,
10
+ 13,
11
+ 16,
12
+ 19
13
+ ],
14
+ "contract_sha256": "5e61fd43ffefc6ccf6882f388b3cec9990c38187c61f836241c4a8bcb9d464f5",
15
+ "cp_size": 1,
16
+ "dmd_models_loaded": false,
17
+ "ema": {
18
+ "checkpoint_decay": 0.9,
19
+ "policy_max_decay": 0.9,
20
+ "policy_min_decay": 0.2,
21
+ "policy_step_decay": 0.01,
22
+ "state_required_on_resume": true
23
+ },
24
+ "evaluation_checkpoint": "model_ema.pt",
25
+ "evaluation_requires_frozen_official_refiner": true,
26
+ "frozen_refiner_in_reward": false,
27
+ "geometry_reward_loaded": false,
28
+ "keep_last_complete_checkpoints": 2,
29
+ "model_identity": {
30
+ "checkpoint_sha256": "95f8dae5fad465a2917b77144f6cbf43a544683c113415b4579885a165a70c2f",
31
+ "config_sha256": "fce36e587f200d67c3acfe7b7c336e36b456e271fa01248dce805bd9c49fe25e",
32
+ "model_id": "Efficient-Large-Model/SANA-WM_streaming",
33
+ "revision": "fb4ef609f5da47d5815e850a5ecec9a12f069b57"
34
+ },
35
+ "native_schedule": [
36
+ 1000,
37
+ 960,
38
+ 889,
39
+ 727,
40
+ 0
41
+ ],
42
+ "num_latent_frames": 19,
43
+ "optimizer": {
44
+ "adamw_betas": [
45
+ 0.9,
46
+ 0.999
47
+ ],
48
+ "adamw_eps": 1e-08,
49
+ "lr": 1e-05,
50
+ "momentum": 0.95,
51
+ "name": "worldcompass_muon",
52
+ "nesterov": true,
53
+ "ns_steps": 5,
54
+ "weight_decay": 0.0001
55
+ },
56
+ "optimizer_updates_per_outer_step": 3,
57
+ "ordered_reward_loaded": false,
58
+ "port_semantics": "post-loss-fix",
59
+ "profile": "released_code_1588e133_sana_native_smoke",
60
+ "required_loss_fix_commit": "8eec3fec45ed5c6e560c25f48db97c8d5efa2e3e",
61
+ "reward": {
62
+ "action_type": "fine_action",
63
+ "action_weight": 2.0,
64
+ "advantage_clip": 2.0,
65
+ "quality_prompt": "A high-quality, ultra-detailed and well-structured image.",
66
+ "rotation_signed_permutation": [
67
+ 1,
68
+ 2,
69
+ 3
70
+ ],
71
+ "rotation_threshold_degrees": 0.2,
72
+ "standard_deviation_scope": "global",
73
+ "target_frame_offset": 3,
74
+ "target_frame_stride": 4,
75
+ "translation_signed_permutation": [
76
+ 1,
77
+ 2,
78
+ 3
79
+ ],
80
+ "translation_thresholds": [
81
+ 0.002,
82
+ 0.005,
83
+ 0.01
84
+ ],
85
+ "visual_type": "hpsv3_generic_quality_drift",
86
+ "visual_weight": 1.0,
87
+ "zero_variance_advantage": 0.0
88
+ },
89
+ "reward_decode_domain": "stage1_causal_vae_before_frozen_refiner",
90
+ "rollout_integrator": "deterministic_euler_native_velocity",
91
+ "rollout_noise_draws_per_candidate_chunk": 1,
92
+ "rollout_transition_renoise": false,
93
+ "save_optimizer_steps": 3,
94
+ "schema_version": "sana_worldcompass_released_port/v1",
95
+ "selected_bottom": 3,
96
+ "selected_top": 3,
97
+ "student_only": true,
98
+ "target_optimizer_steps": 3,
99
+ "target_outer_steps": 1,
100
+ "target_schedule": "outer_step_mod_6",
101
+ "train_timesteps_per_candidate": 2,
102
+ "upstream_commit": "1588e1336e842b03b0a7860c654ebd7c46bb065e",
103
+ "upstream_repository": "https://github.com/Tencent-Hunyuan/HY-WorldPlay",
104
+ "world_size": 2
105
+ }