pretty_name: WorldRide research artifacts
license: apache-2.0
tags:
- world-model
- video-generation
- knowledge-distillation
- reward-forcing
WorldRide model artifacts
This private dataset repository stores the released checkpoints and provenance for Rewarded World-Model Distillation. It does not contain the training datasets.
Released artifacts
| Directory | Purpose | Step | Main file | SHA256 |
|---|---|---|---|---|
ordered_redmd_step1000 |
intermediate ordered action + geometry Re-DMD student | 1000 | model.pt |
ac5e8a3e99ba4f024d225dd34f34f2c26b3faed3fee08ea4a9ac07d998233776 |
ordered_redmd_step2000 |
intermediate ordered action + geometry Re-DMD student | 2000 | model.pt |
d27ee4f4966e27d79a93bfaf635c6f7fbb18e74f499fc83458733564573808b3 |
ordered_redmd_step3000 |
final ordered action + geometry Re-DMD student | 3000 | generator.safetensors |
7ed0112773727613bf21dc477e57c831f427e5774bd6e93f37844f6d954b0f61 |
ordered_redmd_step3000 |
full final training container (generator + critic) | 3000 | model.pt |
95bff3c1287cc670cbdc56c3ebd1aee6f4bff67f02ac279ee0611e65bf0e7312 |
ode_init_step3000 |
exact causal ODE initialization used by the final run | 3000 | model.pt |
a5fbd37d21194f972b487b1191dcad59fa68bb43be889847efa7c940f1160a5e |
ordered_redmd_no_geometry_step1000 |
intermediate action-only Ordered Re-DMD ablation | 1000 | model.pt |
8109a5712a5f0ec0b9527a57c5ed72c84e4702de9de9ddc723e20089dba78b4f |
ordered_redmd_no_geometry_step2000 |
intermediate action-only Ordered Re-DMD ablation | 2000 | model.pt |
28336d5771a70dfb541e2a24c552b4ded24e664720ba11ffdb6fe0f493566da9 |
ordered_redmd_no_geometry_step3000 |
final action-only Ordered Re-DMD ablation | 3000 | generator.safetensors |
ebd22ca83eaf9ec789869f603e145991f52ef26331e8a658c4c6aec35e637e35 |
ordered_redmd_no_geometry_step3000 |
full ablation training container (generator + critic) | 3000 | model.pt |
b9f220138d4b43b41c31cf3820d4d48b69b988c99c54c1f5b744b1f256522118 |
ordered_redmd_no_action_geometry_step1000 |
intermediate equal-weight multi-anchor DMD ablation | 1000 | model.pt |
9e3d123b1da6bfd1465c451f7572a4feeae8d273931377df071d6e63a390b039 |
ordered_redmd_no_action_geometry_step2000 |
intermediate equal-weight multi-anchor DMD ablation | 2000 | model.pt |
dd8f425034b5b6840c5f82fc78edb9eefb7bbe7c0257d124eafb713da2291aa1 |
ordered_redmd_no_action_geometry_step3000 |
final equal-weight multi-anchor DMD ablation | 3000 | generator.safetensors |
567a5c36232cc97e10ea1f2f5e99a0d6f50079e5de21a95c19abe9394f09c21f |
ordered_redmd_no_action_geometry_step3000 |
full ablation training container (generator + critic) | 3000 | model.pt |
1de8a5e8b2e075e71879ed38cd85d5bcd3d53bae1fb1b75b03b90c225d11f897 |
The final student uses four Stage-1 updates with the frozen timestep schedule
[1000, 967, 908, 764, 0]. It is based on the SANA-WM streaming architecture
and requires the corresponding official causal VAE, Gemma text encoder, and
configuration assets. The optional official refiner remains a separate second
stage and is not embedded in these files.
Inference loading
The preferred inference artifact is the generator-only safetensors file:
from safetensors.torch import load_file
state_dict = load_file("ordered_redmd_step3000/generator.safetensors")
missing, unexpected = generator.load_state_dict(state_dict, strict=True)
assert not missing and not unexpected
generator here is the SANA-WM streaming generator wrapper constructed from
the released resolved config. The export contains exactly 872 bfloat16 tensors
and was compared tensor-by-tensor with model.pt["generator"] before release.
For training-state inspection or continuation, load the full container:
import torch
checkpoint = torch.load(
"ordered_redmd_step3000/model.pt",
map_location="cpu",
weights_only=True,
)
generator.load_state_dict(checkpoint["generator"], strict=True)
assert checkpoint["step"] == 3000
assert checkpoint["denoising_step_list"] == [1000, 967, 908, 764, 0]
The full final container also includes a DMD critic; it is not required for
inference. The ODE file is the direct initialization ancestor recorded by the
final run's parent_ode_import.json, not the older legacy step-8000 artifact.
The step-1000 and step-2000 files are full generator-plus-critic snapshots from
the same continuous ordered action + geometry Re-DMD lineage as step-3000;
they are neither ODE initialization nor action-only checkpoints.
The ordered_redmd_no_geometry_step* lineage is the paired w/o Geometry
ablation: action reward and all typed multi-anchor/DMD training budgets are
retained, while geometry is disabled with zero weight. It resumes the same
sealed step-10 action-smoke transaction as the full method. The final receipt
records non-uniform action-derived local-group weights and verifies that every
geometry diagnostic remained zero.
The ordered_redmd_no_action_geometry_step* lineage is the paired w/o Action + Geometry ablation. It retains HPS feasibility measurement, three typed anchors, candidate branching, DMD replay, and the full multi-anchor compute budget, but both semantic reward channels are disabled. Therefore every detached local-group reward weight is exactly one; HPS is diagnostic and cannot create a preference by itself. This is an equal-weight multi-anchor DMD control, not a zero-cost vanilla-DMD topology.
For a strictly paired continuation, this run resumes the shared action-smoke checkpoint at step 10 (model.pt SHA256 b7defa292ab90770a4ee3998f1c3807c204c1fd8ce973a7afa0761f5a77bb549) together with both optimizer states and all RNG states. Formal training uses start_step=10 and max_steps=3000, so its reported global step includes 10 action-rewarded smoke updates; it is not a from-scratch pure equal-weight run. The portable final receipt preserves the historical source-stage label geometry_final, even though its method variant is no_action_geometry. The included calibration measurements document shared lineage but do not produce non-unit reward weights in this ablation.
Provenance and verification
- final code commit:
72bf6df662df8ff19e8aecc4488b81941b666b16 - ODE code commit:
eaaa19d561f6bf368dae7669a296c40ba1336ec1 - SANA-WM model-config SHA256:
fce36e587f200d67c3acfe7b7c336e36b456e271fa01248dce805bd9c49fe25e - parent specification SHA256:
0ff66273574d4cfeca98fba2e0b0f3dfa8bc5af96faaa181a5c56e4241b5834d - no-geometry ablation code commit:
1aa3d986769fc882a9063d31a0f03081ac4a74f5 - no-geometry source receipt SHA256 (before portable path sanitization):
3ac69a48d04a710eaa2a9a021cfc204754fd4bc721cdacb8fa091abbda7410c7 - uploaded portable receipt SHA256:
db7bb51adc67530d468d4f246a2478daf77d08471fce865104e37f121c942911 - no-action/no-geometry ablation code commit:
34119901287f3ebdf61096696191a46feafd653c - no-action/no-geometry common step-10 model SHA256:
b7defa292ab90770a4ee3998f1c3807c204c1fd8ce973a7afa0761f5a77bb549 - no-action/no-geometry source receipt SHA256 (before portable path sanitization):
3001545d48837d52d30f1a9c4ce463dbf28408e64dfecb6ca35b902c294d7f69 - no-action/no-geometry uploaded portable receipt SHA256:
e2159b9b9996f0cdf55cbdc76e961549f8db1400f0950664d58514bf91d4dc2a
metadata/manifest.json describes the selection policy. Every uploaded file is
bound by metadata/checksums.sha256; the generator export has an additional
tensor-equality receipt beside it. Machine-local paths in metadata are replaced
with portable placeholders.
The repository code is Apache-2.0. Users must also follow the licenses and access terms of the upstream SANA-WM, text-encoder, VAE, refiner, and evaluation assets used with these checkpoints.