01AUTHOR REPORTED
RTX 4090 · ComfyUI port
10 seconds of 1280 × 704 video in 85 seconds; 832 × 480 in 34 seconds. Warm run, two 5-second segments.
Port author’s README · checked 2026-10-29
FIELD GUIDE / 02
Hardware, without
the guesswork.
Published configurations and measurements, with the runtime and source in view. No compatibility promises.
01PUBLISHED REPORTS
What has been reported.
The following are the port author’s measurements, not measurements by this site. Results depend on the machine, settings and version.
02AUTHOR REPORTED
12 / 16 / 24 GB · simulated
93 / 91 / 89 seconds respectively for 10 seconds at 1280 × 704. Simulated on an RTX 4090 by holding back the rest of its memory, before the last kernel fusions. Not measured on a 12, 16 or 24 GB card.
Port author’s README · checked 2026-10-29
03AUTHOR REPORTED
RTX 4090 · 20-second world
About 3 minutes 15 seconds at 1280 × 704, including 91 seconds of decoding. Starting from an image adds 2–6 seconds, according to the port author.
Port author’s README · checked 2026-10-29
The simulated 12, 16 and 24 GB figures were taken before the last kernel fusions, according to the author. They do not establish performance or compatibility on actual cards with those capacities.
02SYSTEM MEMORY
VRAM is only part of it.
The port author lists 32 GB of system RAM as a minimum, with more for longer worlds. History beyond the last 18 latent frames stays in system RAM.
~5.9 GBper 5-second segment at 1280 × 704
~2.3 GBper 5-second segment at 832 × 480
~27.6 GBmodel files alone, excluding cache and outputs
Reported by the port author; model file total calculated from published sizes. Hardware requirements .
↳DEMO & LOCAL
Why the demo does not predict your speed.
The official demo runs remotely. The ComfyUI port is a separate implementation with its own memory handling, so demo speed is not a benchmark for your GPU.
The port author says history beyond the latest 18 latent frames stays in system RAM. The low-memory figures above were simulated on an RTX 4090, rather than measured on separate low-memory cards.
Start From Image is an addition by the port author: the port holds the still for 3 latent frames and encodes it into the initial world history. The author says the underlying model has no native image input.
Source: Port capabilities ↗ · checked 2026-10-29
Details: Start From Image ↗ · Camera control guide →
03READ WITH CARE
What we haven’t verified.
As of 2026-10-29, we have not independently verified runs on actual 12 GB or 16 GB cards, 48 GB cards, multi-GPU setups, AMD or Apple hardware. Lack of a report is not proof that a configuration cannot work.
The official reference runtime and this ComfyUI port are separate implementations. Do not assume their memory settings or performance are interchangeable.
Does the 12 GB simulation mean my card will work?
No. The author restricted the memory available on an RTX 4090; that is not a test of another GPU’s hardware or drivers.
Why does system RAM matter?
According to the port author, history beyond the recent window is kept in system RAM and grows as the world gets longer.
Where can I check the latest requirements?
Start with the port’s requirements and speed and memory notes .