File size: 2,366 Bytes
4e282cc | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 | {
"name": "Astronex-World-5B",
"description": "Causal camera- and action-controllable video world model, Wan2.2-TI2V-5B backbone.",
"architecture": "CausalWanModel",
"backbone_variant": "wan2.2",
"parameters": 5351000000,
"dtype": "bfloat16",
"latent": {
"channels": 48,
"compression": [
4,
16,
16
],
"vae": "Wan2.2 VAE"
},
"model_kwargs": {
"model_name": "Wan2.2-TI2V-5B",
"timestep_shift": 5.0,
"local_attn_size": 20,
"sink_size": 4,
"use_camera": true,
"use_action": true,
"action_dim": 64,
"action_output": false,
"num_embodiments": 32,
"prope_scale": 1.0,
"prope_chunk_anchor": true,
"prope_num_layers": 30,
"prope_exclude_sink": false,
"prope_sink_dual": false,
"sink_attention_repeat": 1,
"history_cross_attention": false,
"is_causal": true
},
"inference": {
"sampler": "UniPC",
"sampling_steps": 8,
"num_frame_per_block": 8,
"guidance_scale": 3.0,
"timestep_shift": 10.0,
"cond_prefill_timestep": 100,
"ref_cond_scale": 1.1,
"ref_sequence_conditioning": true,
"rgb_chroma_stabilize_strength": 1.0,
"rgb_chroma_stabilize_fade": 4,
"retrieval_window": true,
"retrieval_recent_frames": 4,
"retrieval_min_overlap": 0.3,
"retrieval_swap_per_block": 2,
"note": "8 steps is the validated setting. 4 steps degrades visibly -- cabinets and signage smear -- so it is not a supported mode."
},
"provenance": {
"base": "logs/merged/online_unipc25_to12_step20_full.pt",
"adapter": "logs/wan22_dmd_mp2/checkpoint_model_001545/model.pt (generator_ema, 600 LoRA tensors)",
"merge": "W = base.weight + (lora_b @ lora_a) * 128/64, folded in fp32 then cast to bf16",
"training": "bidirectional teacher -> causal student via DMD, 1545 steps, with a motion-preservation term (weight 2.0) and rolling forcing at p=0.5"
},
"action_output_head": {
"present": true,
"tensors": 6,
"shape": "LayerNorm(3072) -> Linear(3072, 3072) -> SiLU -> Linear(3072, 64)",
"initialisation": "output layer zero-initialised, so installing it leaves the model unchanged",
"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"
},
"license": "apache-2.0"
}
|