What this is
Code World Model drives a video model with a proxy: a per-frame
stream of metric depth and semantic-class maps that tells the model where everything is.
The released code runs inference on proxies; nothing released builds them. So we built one —
a scripted world simulator and CPU ray tracer, NumPy only, no weights, no GPU.
It produces proxy conditions in the released format, at the
released camera intrinsics — which MiniMax-H3 accepts unmodified.
It does not reconstruct the authors' scenes. We built an
inverse renderer to find out: their water surface is exactly a plane and their road is a
plane to 1.4%, but everything else is film-grade mesh. On a stationary frame, re-rendering
their own recovered point cloud costs 0.03% depth error; our fitted primitives cost 0.60%
— about 20× worse. That is the honest size of the gap.
Six released examples, driven both ways
For each example we author the scenario in our proxy from the authors' own
prompt, then generate video twice — once from their released conditions,
once from ours — holding the anchor image, the caption and the seed fixed. The proxy is the
only variable between the two clips.
ex18 — wading a canal
Chosen for water: class 2 covers 52.8% of the frame, the strongest water
case in the 40. The clearest evidence the proxy is doing work — the figure sits centre-left
in their DUV and centre-left in the video it drives; right-of-centre in ours, and
right-of-centre in the video ours drives.
their proxy
our proxy
the video theirs drives
the video ours drives
ex36 — wide composition
Road 46.9%, sky 26.0% — far wider than our own takes tend to frame.
Near-identical: surveyor on a wet road, water tower, two vehicles.
their proxy
our proxy
the video theirs drives
the video ours drives
ex31 — an animal protagonist
The protagonist is a white stag. It is semantic class 8,
not 9 — which is how we learned that class 8 encodes the protagonist, not
"a human". Both produce the stag; ours sits higher in frame, as our DUV places it.
their proxy
our proxy
the video theirs drives
the video ours drives
ex28 — a vehicle at close range
Class 10 covers 15.2%, the largest vehicle presence in any window 0.
Both render the giant silk-moth across the cobbled lane with the cart at upper right —
although our vehicles are wireframe edges where theirs are solid masses.
their proxy
our proxy
the video theirs drives
the video ours drives
ex30 — a plain walk
Their own prompt: "No vehicle, mount, carried object … is present."
A baseline with nothing to place. Comparable; our vegetation is blobbier and our horizon
sits higher.
their proxy
our proxy
the video theirs drives
the video ours drives
ex34 — flight
Their proxy is a high oblique vista over terrain, a lake and winding roads.
Ours is a ground-level corridor. The two disagree completely — and the videos come
out nearly identical. The anchor and the caption carried this shot, not the proxy.
We keep this panel because six hand-picked successes would misrepresent the method: on some
shots the video model is robust enough that the proxy contributes little. The underlying gap
is real — a vista shot frames terrain, and our rig is a subject-following boom with no such
mode.
their proxy
our proxy
the video theirs drives
the video ours drives
Try it
git clone https://github.com/guangyliu/cwm-proxy && cd cwm-proxy
pip install numpy pillow
python take_actions.py out/my_take # ~25 s, 484 frames of conditions + captions
python validate_conditions.py out/my_take/conditions
Then drive MiniMax-H3 with them:
python portb/run_cp.py --conditions out/my_take/conditions \
--prompt-dir out/my_take/prompts --anchor first_frame.png \
--windows 5 --steps 20 --tag my_take