FeaturesTimeMachineModels BenchmarksPricingDocsGet a key
Local runtime · no cloud rendering

Run 30B-class video on the GPU you already own.

Veizik is a ComfyUI drop-in powered by the LimML engine: low-VRAM capacity-compute + native CUDA DiT. TimeMachine Render branches from a saved checkpoint — keep the approved prefix, re-render only the failing tail.

Free to install & verify Windows (WSL2) / Linux RTX 3090 / 4090 · 24 GB
30B
video model on a 24 GB GPU, no OOM
9.55 GB
LTX-13B peak VRAM (measured)
6/6
native DiT engines bit-exact
50%
GPU saved on a mid-point branch
807 KB
branchable checkpoint timeline
~109s
LTX-13B warm render
What it does

Six things that make big models run on small hardware

The value is execution quality — running the biggest open models reliably on hardware you already own — not license protection.

Low-VRAM execution

Capacity-compute tiering (hot / warm / cold placement) fits 30B-class video on a single 24 GB GPU without OOM.

measured on RTX 3090

Native CUDA DiT

Flagship models run on Veizik's own DiT kernels — N-layer streaming + fp16 tensor-core, verified bit-exact vs reference.

6/6 verified

Weight-resident speed

No per-step reloading; fp16 tensor-core paths give up to ~10× vs fp32 at the layer/kernel level.

profile-conditional

Crash-resume

Step-level checkpoints + retry/degrade + isolation — one job's fault never discards the render or takes down the farm.

tested

TimeMachine Render

Branch from a saved checkpoint: keep the good prefix, re-render only the failing tail with new prompt / style / CFG.

beta · GPU-validated

ComfyUI drop-in

Swap comfyveizik. Same graph, low-VRAM + native DiT. Unknown models fall back — the swap never fails.

run / serve / t2v
Signature feature

Branch instead of re-render

The biggest waste in AI video is "the first part is good but the rest breaks, so I re-render everything." TimeMachine fixes that — one checkpoint infrastructure serves both crash-resume (same settings) and creative branching (new settings).

base render · denoise steps 5 10 15 ◀ branch 20 30 prefix 0..15 — REUSED (kept) tail 16..30 — new style/cfg real GPU (LTX): re-ran 8/16 steps · 50% saved timeline stored: 807 KB · sparse latents only
  • Keep the good part. Composition, faces, product placement, lighting are preserved — not thrown away.
  • Fix only what's wrong. A clip that collapses after ~6 s regenerates only its tail — a fraction of the work.
  • A/B fanout. Several endings (warm / cyberpunk / luxury) from one branch point, in parallel across the farm.
  • Auto-find "from where". A collapse detector flags face drift, flicker, and motion jumps and suggests the branch point.
  • Low storage. Sparse latent checkpoints — a branchable timeline is ~kilobytes, not a copy of every frame.
Compatibility

The best open models — native or universal

Flagship families run on native CUDA DiT (verified bit-exact at the block/engine level). Everything else runs a universal fallback, so a comfy → veizik swap is never rejected.

Native DiT

Step-Video 30B

rel 2.35e-6 · single 24 GB
Native DiT

HunyuanVideo

rel 3.6e-6 · dual/single
Native DiT

Wan 2.1

rel 5.53e-7
Native DiT

LTX-Video

rel 1.94e-7 · measured
Native DiT

CogVideoX

rel 8.43e-7
Native DiT

FLUX.1-dev

rel 4.58e-7 · image
OpenAI-compat

Llama / Qwen

LLM · drop-in endpoint
Universal

Any ComfyUI model

T2 fallback · always runs
Works with your stack
runs on the tools you already use — no rewrite, no lock-in
drop-in

ComfyUI

run / serve / nodes
runtime

NVIDIA CUDA

sm_80+ · fp16 TC
pipeline

diffusers

T2 fallback path
framework

PyTorch

2.x · bf16 / fp16
OS

WSL2 / Linux

Windows or native
API

OpenAI-compatible

LLM drop-in endpoint
Where we are

Measured on real hardware

Numbers are profile-conditional (resolution, frames, steps, model). VRAM is reported from framework allocators — authoritative on WSL2.

Run (RTX 3090, 24 GB)TimePeak VRAMOOM
LTX-13B · 1280×704 · 49f · 40 steps~109 s9.55 GBnone
Step-Video 30B · 992×544 · 51fsingle GPU≤24 GBnone
TimeMachine branch · step 8 of 168/16 steps807 KB tl50% saved
Native DiT engineNumerical match (rel)
Step-Video 30B2.35e-6bit-exact
LTX-Video1.94e-7bit-exact
CogVideoX8.43e-7bit-exact
HunyuanVideo3.6e-6bit-exact
Wan 2.15.53e-7bit-exact
FLUX.1-dev4.58e-7bit-exact
VRAM budget · single RTX 3090 (24 GB)LTX-13B peak 9.55 GB · 40% used
9.55 GB 24 GB
Headroom left on the same card for a 30B-class model — no OOM, no offload penalty on the hot path.

We don't claim uncrackable protection. The durable advantage is execution quality, optimization depth, update cadence, and benchmarks — not secrecy.

Drop-in in 3 steps

Keep your workflow. Change the command.

1 · Install

Download the runtime for Windows (WSL2) or Linux. Runs without a key — the free path proves it runs on your GPU.

2 · Swap the command

Replace comfy with veizik (or vz) in your existing workflow. Same graph, low-VRAM + native DiT.

3 · Render & branch

Render, then branch the parts that broke — never re-render the whole clip again.

How it compares

Local runtime vs cloud tools

VeizikCloud video (Runway / Pika)Raw ComfyUI
Runs on your own GPU yesno — cloudyes
30B video on 24 GB no OOMn/aOOMs
Checkpoint branch (keep prefix) TimeMachinere-renderre-render
Per-render costyour electricity~$0.05–0.95 / clipyour electricity
Data leaves your machine neveruploadednever
Price$9–59 / mo license~$28–76 / mofree / DIY
Plans

Free to verify. Paid to produce.

Pricing is a local runtime license — not cloud credits. You supply the GPU. Checkout — coming soon · free keys are live today.

FREE
$0
  • Verify on your GPU
  • 720p · 5 s · lightning
  • Watermarked · non-commercial
Get free key
CREATOR · MAIN
$24 /mo
  • Watermark removed
  • Commercial use (individual)
  • Denoise-step branch
  • TimeMachine (beta)
PRO
$59 /mo
  • Automation · batch
  • A/B fanout orchestration
  • 1–2 devices
STUDIO / ENT
Contact
  • Multi-GPU · team seats
  • Private deployment
  • Render-farm orchestration
Contact sales

+ TimeMachine Beta Pack — $12/mo · branch render + collapse detection. Licensing is a server-signed entitlement (the runtime, not the UI, enforces features). Launch pricing target; annual discounts planned.

Quickstart

Drop-in usage

# make veizik your comfy
alias vz='veizik'
comfy() { veizik "$@"; }        # comfy run wf.json now runs veizik

veizik doctor                   # probe GPU + per-model support tier
veizik activate <key>          # bind device, receive entitlement token
# direct native text-to-video (low-VRAM + native DiT)
veizik t2v "a barista pouring latte art, warm cafe" \
  --model ltx --w 1280 --h 704 --frames 49 --steps 40

# or run an existing ComfyUI workflow unchanged
veizik run my_workflow.json
# keep the good prefix, re-render only the tail
veizik base   --job clipA --model ltx --prompt "..." --steps 30
veizik branch --job clipA --from-step 15 \
      --style "luxury commercial" --cfg 4.5 --seed-policy branch

# A/B fanout: 3 looks from one branch point
veizik fanout --job clipA --from-step 20 \
      --styles "warm cinematic,cyberpunk neon,luxury glossy"
# auto-detect where it went bad -> suggested branch points
veizik collapse --job clipA
#  - 3.2s: face drift begins
#  - 4.0s: optical-flow discontinuity
#  - 4.6s: prompt adherence drops

veizik timeline --job clipA     # checkpoints + branch lineage + storage
Questions

FAQ

No. Veizik runs entirely on your own GPU — nothing is uploaded and no cloud render credits are involved. A license is a local runtime license; you supply the hardware.

An NVIDIA RTX 3090 / 4090 (24 GB) or better, a recent driver, and Windows (WSL2) or Linux. 30B-class video was measured on a single 24 GB RTX 3090 with no OOM.

No — Veizik is a strict drop-in. Swap comfyveizik and the same graph runs low-VRAM on the native DiT engine. Models without a native engine fall back to a universal path, so the swap is never rejected.

It branches a render from a saved denoise-step checkpoint: the approved prefix is reused and only the failing tail is regenerated with a new prompt / style / CFG. It's the same checkpoint infrastructure as crash-resume. Real-GPU validated on LTX (a mid-point branch re-ran only half the steps). It's live in the CLI today as a beta feature; the paid Beta Pack checkout is coming soon.

No — Veizik is a proprietary commercial runtime. Installation and evaluation are free; production use requires a license. The engine (LimML) is not distributed as source.

Permission is a server-signed entitlement token, not the key itself. Activation binds to a device fingerprint, and feature gating lives in the runtime executor — not just the UI. You can issue and revoke keys above.

Native CUDA DiT (verified bit-exact) for Step-Video 30B, HunyuanVideo, Wan 2.1, LTX, CogVideoX, and FLUX.1; Llama/Qwen-class LLMs via an OpenAI-compatible endpoint; any other ComfyUI model via the universal fallback.

Live

Get an API key

LIVE

Free keys render watermarked, non-commercial. Your key authorizes via a server-signed entitlement — issued below, enforced in the runtime.

Your API key (shown once — copy it now):
Set once: export VEIZIK_API_KEY=… then veizik doctor.

Keys are stored hashed; the raw key is shown only once. Revoke anytime — the entitlement stops validating.

Roadmap
Coming soon

Being finalized

SOON

TimeMachine Studio

Visual timeline branching — generation filmstrip, collapse auto-detect, one-click A/B fanout in the browser.

SOON

Paid checkout

Creator / Pro / Studio via merchant-of-record (USD, tax handled). Your free account carries over.

SOON

Render Farm

Multi-GPU orchestration, team seats, API, parallel branch candidates with auto-evaluation.

Grab a free key today — the engine, drop-in CLI, and TimeMachine (beta) are live; the surfaces above are on the way.