veizik WAN (CUDA) v1.2.91 — Wan 2.1 & Wan 2.2 video — cuda-wan-1.2.91

Text-to-video and image-to-video with Wan2.1-T2V-1.3B, Wan2.1-T2V-14B, Wan2.2-T2V-A14B and Wan2.2-I2V-A14B on your own NVIDIA GPU. Veizik runs the models through its own native engine: no Python, no CUDA toolkit. You supply the weights; the frames are written as PNG files to the folder you name. It also runs a resident render server with a host-loop (bin/veizik-resident) that keeps one process and one GPU context across requests, and mid-render preemption. A maintenance cut over 1.2.8: the base text- and image-to-video renders are byte-identical (the engine's non-LoRA code is unchanged), distributed now as a single signed .vzpack.

This page records the state as of 2026-10-30 and is not updated as the catalogue moves. For what runs now, see models.

What this build does

• veizik run <model key> "prompt" --out DIR renders a video and writes its frames as PNG files to DIR/frames/; nothing is written inside the package.
• Serve: veizik serve <model> runs a resident render server that renders request after request in one process and one GPU context, with results bit-identical to veizik run.
• Four modes: quality (the reference path and the default), balanced, fast and ultrafast.
• An interrupted render continues from its last step with --resume, and the result is identical to a render that never stopped.
• Every file of the package is checked against its sealed manifest before each run.
• GPU code for sm_80, sm_86, sm_89, sm_90a, sm_100a and sm_120a.
• Every frame it writes carries the provenance mark described in the Terms: an in-pixel watermark plus XMP DigitalSourceType trainedAlgorithmicMedia.
• Distributed as a single signed .vzpack: veizik install wan22 fetches and verifies it (and the shared-core .vzroot once) from its GitHub release and refuses any other file.
• mid-render preemption — POST /v1/videos/generations with "priority": "urgent" interrupts a running render at a step boundary, is served, then the first render resumes from its checkpoint.

Models in this release

Platform

Linux on x86-64

Verification

This build was withdrawn (unrunnable) and is no longer offered. Superseded by cuda-wan-1.2.92.

File

a previous build — not downloadable

Size

7,007,814 bytes

SHA-256

ec21840bb29e63284743c92b5c19ae4bacef7e4cd07ddc17e78192d1b1b1cadb

Verify, install, check

$ shasum -a 256 -c SHA256SUMS
$ ./install.sh
$ veizik doctor

Full steps are in the documentation.

Fixed issues

• Maintenance cut over 1.2.8: the base text- and image-to-video renders are byte-identical (the engine's non-LoRA code is unchanged); carries the 1.2.5/1.2.6/1.2.8 install-reliability and resume improvements forward.
• Internal engine preparation for the distilled image-to-video Lightning path; that model is not yet enabled on this site and is not advertised until its render is verified.

Known issues in cuda-wan-1.2.91

• Model weights are not included. veizik pull wan22 downloads the Wan2.1/2.2 weights (Apache-2.0) and prepares the video decoder automatically — no manual VAE setup; veizik pull --prep-local DIR prepares weights you already have.
• Speed figures for this package were measured on an NVIDIA RTX 3090 (sm_86) only.
• The image-to-video model takes its input image as raw RGB bytes, one byte per channel, at exactly the render size (width × height × 3 bytes).
• balanced, fast and ultrafast are not bit-identical to quality; without --det-vae yes decoded frames may differ slightly from run to run.
• The distilled image-to-video Lightning model is not yet enabled: its render is being verified before it is advertised.