The state layer for AI
by VoyceMe

Everything changes.
Crıcket keeps what’s true now.

Crıcket is the state layer for AI. Our API layer maintains what’s true in a scene, so the next output keeps it.

Works acrossVideoLLMsImagesRoboticsWorld models
Work with us
Wan 3.0 aloneChanged
Wan 3.0 + CrıcketSame
Shot 01 of 06 · same prompt, same model
Smoke test Wan 3.0Real output, unedited

Same prompts. One stays consistent.

Six one-line prompts. Wan alone changes the woman and the dog. With Crıcket, they stay the same in every shot.

Wan 3.0 aloneChanges every shot
Wan 3.0 + CrıcketSame in all six
    6/6shots keep the same woman and dog
    1 lineper prompt, nothing repeated
    0changes to Wan

    This is our first video test, on a closed model. It’s about 30–40% of what Crıcket can do. Tap any frame to play both clips.

    How it works

    You write the line.
    Crıcket keeps the scene.

    Follow one shot, shot 04, through the system.

    1

    You write

    PromptShot 04
    They cross the old stone bridge.
    Generate

    One short line. No character details.

    2

    Crıcket maintains

    • Carried into shot 04
    • Lin: red raincoat, yellow boots
    • Mochi: corgi, blue collar
    • The skewer from shot 02
    • Rainy night market

    What’s already true from earlier shots goes into the new one.

    3

    Wan renders

    Shot 4 with CricketMatches shot 03Shot 4 without CricketWan alone

    Same woman, same dog, now on the bridge. Without Crıcket, Wan loses them.

    Working together

    Crıcket on every Alibaba video and image model.

    Alibaba has more than a dozen video and image models. Open source or closed, Crıcket sits on top of any of them.

    Alibaba models
    Wan videoWan imageQwen-Image

    maintains what’s true

    Long videos. Consistent images.

    Same character, same world, every shot and every frame.

    Video

    Wan 3.0 · 3.0 Prime
    Wan 2.7
    Wan 2.6 · 2.5
    Wan 2.2
    Wan 2.1
    Wan-Animate-2 · Wan-Dancer

    Imagery

    Qwen-Image 3.0 · 3.0 Pro
    Qwen-Image 2.1 Pro
    Wan 2.7 Image · Image Pro
    Wan 2.6 Image
    Qwen-Image · Qwen-Image-Edit

    Any way you prompt

    Seasoned or brand new.
    Crıcket makes it consistent.

    Crıcket takes any style of prompting. Confident Wan users who write detailed prompts, or people writing their very first one. Either way, the output stays consistent.

    Confident Wan users

    “Lin, a woman in a red raincoat and yellow boots, and Mochi, an orange corgi with a blue collar, cross the old stone bridge.”Detailed, written the Wan way
    Crıcket keeps every detail from drifting

    Even expert prompts drift from shot to shot. Crıcket holds the faces and the world in place.

    See detailed prompts

    New to prompting

    “They cross the old stone bridge.”One plain line
    Crıcket fills in the rest

    No prompting skill needed. Crıcket turns one line into a consistent scene.

    See simple prompts

    Two ways in

    Two ways to work together.

    Recommended

    License the source

    Our team builds a version of Crıcket just for Wan.

    • Custom built for Wan’s video
    • Higher performance
    • Lower cost
    Also possible

    Per API call

    Plug Crıcket into any Wan call as it is. Not custom built, and it costs more.

    Exclusive for video is open. We’re already talking with a few others.

    Let’s talk

    Let’s build it
    together.

    We have interest across many verticals. Video is the one we want to build with Alibaba.

    Email Dylan Telano

    telano@voyce.me · Built by VoyceMe, the largest English-original manga and webtoon studio in North America.

    Shot 01

    Wan 3.0 alone
    Wan 3.0 + Crıcket
    Same one-line prompt, same model.