veizik FLUX.1 (CUDA) v1.0.5 — dynamic resolution — cuda-flux-1.0.5

Text-to-image with FLUX.1 on your own NVIDIA GPU — eight variants: FLUX.1-dev (12B) and FLUX.1-schnell, plus Kontext, Krea and the Tools set (Fill, Depth, Canny, Redux). Veizik runs the models through its own native engine: no Python, no CUDA toolkit. You supply the weights under their Black Forest Labs licences; images are written as PNG at the resolution you ask for — the default is 1024x1024, and this build adds free --w/--h (a multiple of 16, in the range 256-3072), so portrait like --w=768 --h=1344 and other sizes render too. Distributed as a single signed .vzpack, and render-verified fresh on the shipped bytes.

This page records the state as of 2026-11-08 and is not updated as the catalogue moves. For what runs now, see models.

What this build does

• veizik run flux1dev "prompt" --out out.png renders to the file you name at the default 1024x1024, or any size you ask for with --w/--h (a multiple of 16, range 256-3072) including portrait like --w=768 --h=1344; the eight keys are flux1dev, fluxschnell, fluxkrea, fluxcanny, fluxdepth, fluxfill, fluxkontext and fluxredux.
• Seven modes: quality (the reference path), balanced, efficient, compact, low-memory, fast and ultrafast. The Tools variants condition on an input image via --image (Fill also takes --mask).
• Per-architecture cubin for sm_80, sm_86, sm_89, sm_90a, sm_100a and sm_120a — no compile step at render time.
• Every file of the package is checked against its sealed manifest before each run.
• Every image 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 flux1dev fetches and verifies it (and the shared-core .vzroot once) from its GitHub release and refuses any other file.

Models in this release

Platform

Linux on x86-64

Verification

The file name, size and SHA-256 are printed on this page and on the release page, and the two agree. It ships as a signed single-file package (.vzpack) listed in SHA256SUMS on the release, and veizik install checks the SHA-256 before the package is used. Every image the package generates carries a v1 provenance mark.

File

veizik-flux-1.0.5-linux-x86_64.vzpack

Size

45,274,050 bytes

SHA-256

33921b9c0d6690fff86f53be87583cc1c02ad5ddacfa4af4b2e06b350da1a62a

GPU architectures

sm_80, sm_86, sm_89, sm_90a, sm_100a, sm_120a

Release

v1.0.5

Get, install, activate

# the veizik command for Linux x86-64
$ curl -fsSL https://veizik.com/install.sh | sh
# if the installer asks you to put ~/.local/bin on your PATH
$ export PATH="$HOME/.local/bin:$PATH"
# accepted names: flux or a full model key (flux1dev, fluxschnell, fluxkrea, fluxcanny, fluxdepth, fluxfill, fluxkontext, fluxredux); any case
$ veizik install flux
$ veizik activate
# without a browser (SSH, servers): create a key + CLI password on your account page, then:
$ veizik activate --key KEY --password-stdin

veizik install refuses a download whose SHA-256 differs from this release. The other packages are on the download page.

Fixed issues

• Output resolution is now a runtime option: --w and --h (each a multiple of 16, in the range 256-3072) render any size including portrait, where earlier builds were fixed at 1024x1024. The default is still 1024x1024, and that default output is bit-for-bit identical to v1.0.4 — the engine and every GPU kernel are unchanged; only the pipeline driver was re-cut to read --w/--h.

Known issues in cuda-flux-1.0.5

• Model weights are not included, and two licences apply. FLUX.1-schnell is Apache-2.0 with no conditions; the dev variants (dev, Kontext, Krea, Fill, Depth, Canny, Redux) are FLUX.1 [dev] under the FLUX.1 Non-Commercial Licence — accept it at Black Forest Labs, then veizik pull <key> downloads the weights with your Hugging Face token; veizik run … --weights DIR uses weights you already have.
• Commercial use of the FLUX.1 [dev] variants requires a separate licence from Black Forest Labs. Veizik neither redistributes nor relicenses the weights.
• balanced, efficient, compact, low-memory, fast and ultrafast are not bit-identical to quality.