Skip to content

docs(hy_worldplay): add HY-WorldPlay WAN-5B I2V model card#336

Open
wenqingw-nv wants to merge 4 commits into
mainfrom
dev/wenqingw-nv/hy-worldplay-model-card
Open

docs(hy_worldplay): add HY-WorldPlay WAN-5B I2V model card#336
wenqingw-nv wants to merge 4 commits into
mainfrom
dev/wenqingw-nv/hy-worldplay-model-card

Conversation

@wenqingw-nv

@wenqingw-nv wenqingw-nv commented Jun 13, 2026

Copy link
Copy Markdown
Collaborator

Adds the HY-WorldPlay WAN-5B I2V model card under docs/source/models/, following the LingBot-World card format. Videos are generated from HY-worldplay's provided images.

Contents

  • docs/source/models/hy_worldplay.rst — overview, requirements, install, run commands, hero + sample-video grid, profiling benchmark, citation.
  • docs/source/models/index.rst — adds HY-WorldPlay to the model list.
  • docs/source/index.rst — adds HY-WorldPlay to the Models nav toctree (so the card shows in the sidebar).
  • docs/source/_static/performance/hy_worldplay/perf-0530.md — benchmark data for the profiling chart.

Sample videos

Native 8-chunk rollouts (num_chunk=8, pose=w-31, seed 0) for assets/img/{1,2,4,8,9}:

  • Hero (top): img-1
  • Grid: img 2, 4, 8, 9 (each labeled with its prompt)

Note: the <source> URLs point at research.nvidia.com/.../assets/hy_worldplay/hy-worldplay-wan-i2v-5b-{1,2,4,8,9}.mp4. The mp4s still need uploading to that CDN before the players render — files are staged locally (/tmp/hy_card_videos/).


Update — videos regenerated with the VAE checkerboard fix (#338). The original sample videos had a ~2px VAE-decode checkerboard (flashdreams Wan 2.2 VAE patchify channel-order bug). Fixed in #338; the 5 sample mp4s (img 1, 2, 4, 8, 9) were regenerated 8-chunk with that fix and need re-uploading to the research.nvidia.com/.../assets/hy_worldplay/ CDN. Card RST is unchanged (same <source> URLs).

Adds the model card (overview, install, run commands, sample-video grid,
profiling benchmark) for the HY-WorldPlay WAN-5B I2V integration, plus its
index entry and benchmark data. Sample grid shows native 8-chunk rollouts
for assets/img {1,2,4,8,9}.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
@copy-pr-bot

copy-pr-bot Bot commented Jun 13, 2026

Copy link
Copy Markdown

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

@greptile-apps

greptile-apps Bot commented Jun 13, 2026

Copy link
Copy Markdown
Contributor

Greptile Summary

Adds the HY-WorldPlay WAN-5B I2V model card to the FlashDreams docs, following the LingBot-World card format and wiring it into both toctree indexes.

  • hy_worldplay.rst covers installation, the flashdreams-run invocation with distilled-checkpoint and pose flags, a four-video sample grid, and a single-device GB300 profiling benchmark vs. the official implementation.
  • perf-0530.md supplies the benchmark table (official: 1578 ms, FlashDreams: 1015 ms) consumed by the existing benchmark_chart.js widget.
  • Both index.rst files are updated to list HY-WorldPlay after LingBot-World, keeping the two nav trees in sync.

Confidence Score: 5/5

Docs-only change adding a new model card; no code paths are affected.

All four files are purely documentation. The new card is self-contained — the benchmark data file, toctree entries, and model list update are all consistent with each other. The only gap is a missing Requirements section; it does not prevent the docs from building or rendering correctly.

docs/source/models/hy_worldplay.rst — missing hardware Requirements section.

Important Files Changed

Filename Overview
docs/source/models/hy_worldplay.rst New HY-WorldPlay model card; missing Requirements section (VRAM/PyTorch) compared to the LingBot-World template it follows.
docs/source/_static/performance/hy_worldplay/perf-0530.md New benchmark data file with a single GB300 row; straightforward, no issues.
docs/source/index.rst Adds HY-WorldPlay toctree entry after LingBot-World; correct and consistent with models/index.rst placement.
docs/source/models/index.rst Adds HY-WorldPlay bullet to the implemented-models list; no issues.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[docs/source/index.rst\ntoctree] --> B[models/hy_worldplay.rst]
    C[docs/source/models/index.rst\nmodel list] --> B
    B --> D[Installation\nuv sync integrations/hy_worldplay]
    B --> E[Running the method\nflashdreams-run hy-worldplay-wan-i2v-5b]
    B --> F[Sample video grid\nresearch.nvidia.com CDN mp4s]
    B --> G[Profiling benchmark]
    G --> H[_static/performance/hy_worldplay/perf-0530.md\nGB300: official 1578ms / FlashDreams 1015ms]
    H --> I[benchmark_chart.js\nrenders bar chart]
Loading

Reviews (4): Last reviewed commit: "docs(hy_worldplay): drop img-1 from samp..." | Re-trigger Greptile

Comment on lines +22 to +23
<a class="model-link-button" href="https://github.com/Tencent-Hunyuan/HY-WorldPlay" target="_blank" rel="noopener noreferrer">Project page</a>
<a class="model-link-button" href="https://github.com/Tencent-Hunyuan/HY-WorldPlay" target="_blank" rel="noopener noreferrer">Official code</a>

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Duplicate link for "Project page" and "Official code"

Both buttons resolve to https://github.com/Tencent-Hunyuan/HY-WorldPlay. In the reference card (lingbot_world.rst) the two buttons point to distinct destinations — a dedicated project/landing page and the GitHub repo. With identical hrefs, one of the buttons is redundant and visitors clicking "Project page" land on a code repository, not a project overview. If there is no separate project page yet, consider removing one button or updating the href to the Hugging Face model page (https://huggingface.co/tencent/HY-WorldPlay), which is already referenced later in the file.

Comment on lines +145 to +173
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 <https://github.com/Tencent-Hunyuan/HY-WorldPlay>`_
under matched settings.

.. raw:: html

<figure class="benchmark-figure-wrap">
<div
id="hy-worldplay-benchmark-chart"
class="benchmark-figure"
data-benchmark-md-url="../_static/performance/hy_worldplay/perf-0530.md"
data-benchmark-series="official:Official Impl:#3b82f6;flashdreams:FlashDreams:#76B900"
data-chart-aria-label="HY-WorldPlay benchmark chart"
></div>
<figcaption>
<p class="model-footnote">
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
<a href="https://github.com/NVIDIA/flashdreams/tree/main/integrations/hy_worldplay/tests/parity_check">this instruction</a>.
</p>
</figcaption>
</figure>
<script src="../_static/js/benchmark_chart.js"></script>

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 "VAE encode/decode" vs "VAE-decode" inconsistency in benchmark copy

The section heading on line 148 describes "total DiT + VAE encode/decode runtime", but the figcaption on line 164 narrows that to "total DiT + VAE-decode runtime per autoregressive chunk". If VAE encoding is excluded from the measured interval (as the figcaption implies), the section-level description is inaccurate and may mislead readers comparing numbers across models.

Comment on lines +172 to +173
</figure>
<script src="../_static/js/benchmark_chart.js"></script>

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Citation section listed in PR description but absent from file

The PR description's "Contents" bullet explicitly lists "citation" as one of the card's sections, matching the pattern in lingbot_world.rst (which ends with a BibTeX Citation block). The section is missing here. If HY-WorldPlay has a corresponding arXiv/paper reference it should be added; if not, the PR description should be updated to reflect that the section was intentionally omitted.

jmccaffrey-nv and others added 3 commits June 13, 2026 03:11
The model card sidebar nav is driven by the hidden toctree in
docs/source/index.rst, not the bullet list in models/index.rst. Add the
HY-WorldPlay entry there so the card shows in navigation (was building as
an orphan page).

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
The hero slot pointed at a hy-worldplay-hero.mp4 that doesn't exist; point
it at the img-1 native rollout (hy-worldplay-wan-i2v-5b-1.mp4) so the top
slot shows a real sample.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Img-1 is shown in the hero slot, so remove its duplicate grid card; the
grid keeps img 2, 4, 8, 9.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
@wenqingw-nv

Copy link
Copy Markdown
Collaborator Author

/ok to test 07153d7

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants