Upload config.json with huggingface_hub
Browse files- config.json +67 -0
config.json
ADDED
|
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"name": "Astronex-World-5B",
|
| 3 |
+
"description": "Causal camera- and action-controllable video world model, Wan2.2-TI2V-5B backbone.",
|
| 4 |
+
"architecture": "CausalWanModel",
|
| 5 |
+
"backbone_variant": "wan2.2",
|
| 6 |
+
"parameters": 5351000000,
|
| 7 |
+
"dtype": "bfloat16",
|
| 8 |
+
"latent": {
|
| 9 |
+
"channels": 48,
|
| 10 |
+
"compression": [
|
| 11 |
+
4,
|
| 12 |
+
16,
|
| 13 |
+
16
|
| 14 |
+
],
|
| 15 |
+
"vae": "Wan2.2 VAE"
|
| 16 |
+
},
|
| 17 |
+
"model_kwargs": {
|
| 18 |
+
"model_name": "Wan2.2-TI2V-5B",
|
| 19 |
+
"timestep_shift": 5.0,
|
| 20 |
+
"local_attn_size": 20,
|
| 21 |
+
"sink_size": 4,
|
| 22 |
+
"use_camera": true,
|
| 23 |
+
"use_action": true,
|
| 24 |
+
"action_dim": 64,
|
| 25 |
+
"action_output": false,
|
| 26 |
+
"num_embodiments": 32,
|
| 27 |
+
"prope_scale": 1.0,
|
| 28 |
+
"prope_chunk_anchor": true,
|
| 29 |
+
"prope_num_layers": 30,
|
| 30 |
+
"prope_exclude_sink": false,
|
| 31 |
+
"prope_sink_dual": false,
|
| 32 |
+
"sink_attention_repeat": 1,
|
| 33 |
+
"history_cross_attention": false,
|
| 34 |
+
"is_causal": true
|
| 35 |
+
},
|
| 36 |
+
"inference": {
|
| 37 |
+
"sampler": "UniPC",
|
| 38 |
+
"sampling_steps": 8,
|
| 39 |
+
"num_frame_per_block": 8,
|
| 40 |
+
"guidance_scale": 3.0,
|
| 41 |
+
"timestep_shift": 10.0,
|
| 42 |
+
"cond_prefill_timestep": 100,
|
| 43 |
+
"ref_cond_scale": 1.1,
|
| 44 |
+
"ref_sequence_conditioning": true,
|
| 45 |
+
"rgb_chroma_stabilize_strength": 1.0,
|
| 46 |
+
"rgb_chroma_stabilize_fade": 4,
|
| 47 |
+
"retrieval_window": true,
|
| 48 |
+
"retrieval_recent_frames": 4,
|
| 49 |
+
"retrieval_min_overlap": 0.3,
|
| 50 |
+
"retrieval_swap_per_block": 2,
|
| 51 |
+
"note": "8 steps is the validated setting. 4 steps degrades visibly -- cabinets and signage smear -- so it is not a supported mode."
|
| 52 |
+
},
|
| 53 |
+
"provenance": {
|
| 54 |
+
"base": "logs/merged/online_unipc25_to12_step20_full.pt",
|
| 55 |
+
"adapter": "logs/wan22_dmd_mp2/checkpoint_model_001545/model.pt (generator_ema, 600 LoRA tensors)",
|
| 56 |
+
"merge": "W = base.weight + (lora_b @ lora_a) * 128/64, folded in fp32 then cast to bf16",
|
| 57 |
+
"training": "bidirectional teacher -> causal student via DMD, 1545 steps, with a motion-preservation term (weight 2.0) and rolling forcing at p=0.5"
|
| 58 |
+
},
|
| 59 |
+
"action_output_head": {
|
| 60 |
+
"present": true,
|
| 61 |
+
"tensors": 6,
|
| 62 |
+
"shape": "LayerNorm(3072) -> Linear(3072, 3072) -> SiLU -> Linear(3072, 64)",
|
| 63 |
+
"initialisation": "output layer zero-initialised, so installing it leaves the model unchanged",
|
| 64 |
+
"status": "included but untrained: enable with model_kwargs.action_output=true and train it with the action-output post-training branch before reading actions from it"
|
| 65 |
+
},
|
| 66 |
+
"license": "apache-2.0"
|
| 67 |
+
}
|