HY-WorldPlay#

Introduced by Tencent Hunyuan, HY-WorldPlay is a real-time interactive image-to-video (I2V) world model with action + camera-trajectory conditioning and reconstituted-context memory. FlashDreams ships a native port of the distilled WAN-5B variant (Wan 2.2 TI2V-5B backbone, 4-step distilled Euler).

Generated with FlashDreams' native HY-WorldPlay WAN-5B I2V pipeline.

Installation#

# from the repo root
uv sync --project integrations/hy_worldplay

Running the method#

HY-WorldPlay WAN-5B is image-to-video only. Launch it via flashdreams-run, passing a first-frame image (or --example-data) and the distilled checkpoint:

uv run --project integrations/hy_worldplay \
    flashdreams-run \
    hy-worldplay-wan-i2v-5b \
    --example-data True \
    --num-chunk 8 \
    --pose "w-31"

We provide the following variants:

Method

Description

hy-worldplay-wan-i2v-5b

Wan 2.2 TI2V-5B backbone with action + camera conditioning, PRoPE attention, and reconstituted-context memory. Distilled, 4 steps, streaming autoregressive VAE.

To inspect all supported CLI arguments and their default values, run:

uv run --project integrations/hy_worldplay \
    flashdreams-run \
    hy-worldplay-wan-i2v-5b \
    --help

Some generated samples from the above commands:

a person walking
Walking through a seaside village
Walking through a snowy forest
Walking toward a castle

Profiling benchmark#

Here is the profiling benchmark on total DiT + VAE encode/decode runtime for FlashDreams HY-WorldPlay compared to the official HY-WorldPlay implementation under matched settings.

This chart shows total DiT + VAE-decode runtime per autoregressive chunk (4 diffusion steps) in milliseconds, at steady state (median of the post-warmup chunks), measured at num_chunk=8, 704x1280, seed=0 on a single GB300. For an apples-to-apples comparison, both implementations are forced to use the cuDNN attention backend and torch.compile under matched runtime settings. For the official HY-WorldPlay implementation, see this instruction.