File size: 4,213 Bytes
59630ba | 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 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 | {
"method": "LIVE",
"dataset": "RE10K",
"official_repository": "https://github.com/Junchao-cs/LIVE",
"official_commit": "5f0fe658330c64c585874acad3149aa15ba82d1b",
"checkpoint_repository": "junchaoh-cs/LIVE-Re10K",
"checkpoint_files": [
"live-re10k.ckpt",
"vae-kl16.ckpt"
],
"checkpoint_global_step": 20500,
"weight_source": "raw state_dict; verified no EMA keys",
"vae_weight_consistency": "Every VAE inference tensor in final checkpoint equals the separate released VAE tensor exactly",
"backbone_checkpoint_tensors": 174,
"backbone_checkpoint_elements": 772346944,
"backbone_parameters_verified": 772346944,
"vae_checkpoint_tensors": 312,
"vae_checkpoint_elements": 66456019,
"vae_parameters_verified": 66456019,
"optimizer_groups": [],
"training_lr_stages": "not applicable: inference only",
"evaluation_trainable_parameters": 0,
"lora_or_adapters": false,
"activation_checkpointing": false,
"per_gpu_batch_size": 1,
"global_parallelism": "independent per-scene workers, 8 GPUs per Berzelius node",
"sampler": "official 18-step flow DPM, shift 3, one frame per step",
"selector": "native contiguous latest 31 context frames + one generated frame",
"context_noise": true,
"kv_cache": false,
"network_precision": "native fp16 autocast",
"vae_precision": "fp32",
"vae_scale": 0.2325,
"vae_posterior": "native stochastic sample",
"common_protocol": {
"scene_scope": "final100 (monitor8 smoke subset)",
"trajectory": "G0..G199,G199..G0",
"source_stride": 1,
"observed_frames": 100,
"generated_frames": 300,
"resolution": [
256,
256
],
"camera_mapping": "DFoT normalized intrinsics plus raw world-to-camera extrinsics; native window-first normalization",
"targets": "shared DFoT-prepared RGB decoded and preprocessed by shared/protocol.py",
"model_inputs": "observed RGB only; future camera poses allowed, future RGB excluded",
"output": "generated suffix only; shared saver and metric implementation"
},
"adaptations": [
"Use exact shared scene manifest, RGB preprocessing and camera timeline",
"Pass all 100 observed frames to native autoregressive sampler; preserve its 32-frame context limit",
"Disable compilation for portable bounded smoke and evaluation",
"Skip unused native metric models; scoring is external and common across methods",
"Skip unused FlexAttention mask construction when official SDPA is selected",
"Construct deterministic cycle masks on the existing buffer device; normal module device transfer moves them",
"Normalize compiled checkpoint key prefixes and strictly load all inference tensors"
],
"paper_protocol_difference": "Official paper evaluates skip 2 with one prompt frame and open rollout; common WorldMem evaluation uses skip 1, 100 observations and a closed revisit. Published scores cannot be directly compared.",
"required_runtime": {
"python": "3.10",
"torch": "2.6.0+cu124",
"lightning": "2.5.2",
"torchmetrics": "1.7.4",
"diffusers": "0.34.0",
"attention": "torch SDPA efficient_attention (no xformers or FlashAttention dependency)"
},
"verification": {
"official_clone": "passed",
"both_weight_downloads": "passed",
"checkpoint_weight_audit": "passed",
"python_syntax": "passed",
"setup_shell_syntax": "passed",
"environment_install": "passed; exact released requirements plus pytorch-lightning 2.5.2 and setuptools 80.9.0",
"production_model_preflight": "passed on CPU in dedicated env; strict load of all 486 inference tensors",
"real_data_preflight": "passed; first monitor8 scene RGB400x3x256x256, cameras400x16, exact duplicated turnaround and closed endpoint",
"native_sampler_boundary_check": "passed; actual 18-step sampler with a constant-flow probe preserves observed latents and matches five explicit sliding camera windows through reversal",
"local_gpu_smoke": "root agent schedules after preflight"
},
"numerical_failure_policy": "Raise FloatingPointError on sampler/decoder NaN/Inf; never replace with zero frames",
"evaluation_case_seed": "1_000_003 * metadata_index + 7 at protocol seed0; NumPy seeds modulo2**32"
}
|