QVAC SDK 0.22 adds an image-to-video entry point for MiniMax H3: pass a first-frame image to video() with mode: 'img2vid'. This is a local application integration option, rather than a new H3 checkpoint or a hosted video subscription.
The QVAC release announcement dates to October 19, 2026. This guide follows its SDK 0.22.0 release notes and repository quickstart, reviewed October 22. We have not installed QVAC or generated a clip through it. H3Vid does not currently execute QVAC models in the browser or on the server.
What the new mode actually supplies
The release example uses a loaded model identity, an image buffer, a prompt and explicit frame dimensions. The still anchors the clip's opening. This differs from supplying several reference pictures and asking H3 to invent a new scene around their subjects.
QVAC's repository describes video generation through a diffusion backend alongside its other local inference engines. The SDK provides an application-facing interface; the model files, engine configuration and hardware still determine whether a particular request can run.
Do not turn “local” into “works on every laptop.” The announcement's CUDA support explicitly discusses LLM, embedding and VLA models. That sentence alone is not an H3 video compatibility matrix. Before selecting hardware, check the video engine's current supported model bundle and backend instructions.
Check your Node version before debugging H3
The release changes the bundling requirement to Node ^22.21.0 || >=24.9.0. The notes specifically mention earlier Node 22 versions failing inside the bundler's lexer. A working Node installation can therefore still be too old for this SDK release's packaging flow.
Pin the SDK while reproducing these notes:
npm install @qvac/[email protected]
Use the project's package-manager equivalent if needed. Confirm the resolved package version and runtime version before comparing behavior with a release example. The repository also documents a Python client, but a language binding does not remove the native engine and model requirements underneath it.
For a first integration, establish the repository's basic load, run and unload lifecycle in an isolated sample application. Then configure the intended diffusion video model. An LLM quickstart proves that SDK communication works; it does not prove that H3's weights and video engine are ready.
Read the image-to-video example precisely
The following reproduces the documented request shape. It is a fragment that assumes modelId already identifies a correctly loaded H3 video model, not a complete model installation script:
import fs from "node:fs";
import { video } from "@qvac/sdk";
const firstFrame = fs.readFileSync("first-frame.png");
const { outputs } = video({
modelId,
mode: "img2vid",
prompt: "A sailboat glides across a calm bay.",
init_image: firstFrame,
width: 864,
height: 480,
video_frames: 22,
fps: 24,
cfg_scale: 1,
});
init_image is the image bytes in this example. A filesystem pathname or a cloud URL is a different value; do not substitute one without checking the current API's accepted input types. Keep the image readable and the initial scene simple while verifying that the engine receives it.
The example requests 22 frames at 24 fps, less than one second of nominal playback. It is not a five-second production recipe or a latency measurement. Frame count and fps jointly describe playback duration; they do not describe how long generation takes. Follow the current output interface to consume and save outputs rather than assuming the call instantly returns an MP4 file.
Model fit is an estimate with a policy
SDK 0.22 changes assessModelFit to ask the relevant engine's fitter instead of using a shared coefficient table. Its result distinguishes native fit evidence from computed-only evidence. The old estimate fields are removed; floorBytes carries the computed-only bound.
The load policy is now per call. log, the default, performs the probe without blocking. refuse rejects a load the engine projects will not fit. off skips that probe for the individual load. Switching it off does not create memory or make a model compatible.
The release also notes that an engine-level fit can still be judged too large after the interactive reserve is applied. Record the evidence class and resolved load settings when investigating that difference. A resource estimate is useful for rejecting unsuitable configurations early, but it is not proof of a successful video run.
Keep unrelated additions separate
The same release introduces Laya typed decisions, sound-effect models, speech changes and parallel tool calls. These are separate engines and APIs. Their availability does not mean H3 video automatically receives a speech-generation pass, instant voice cloning or a multi-reference mode.
If your application needs a generated soundtrack, inspect the H3 video output contract and available audio behavior directly. If it needs a separately composed effect, QVAC's sound-effect API is another operation with its own model and resource use. Avoid combining both in the first debugging case.
For the next step, save the exact package lock, model identity, engine settings and a minimal still-to-video request. Verify the resulting file's dimensions, frame rate and opening frame before adding a longer narrative. For a node-based local route, see MiniMax H3 ComfyUI setup; for a hosted product workflow, reference-to-video is a separate execution option.
The cover is an imagegen concept illustration of local image-to-video integration, not a QVAC screenshot or a generated H3 sample.