"Released for research use" reads as a limit on what you may do with the weights, and Apache-2.0
imposes no such limit: it grants commercial use to everyone. A notice that appears to restrict
what the licence grants creates ambiguity about which one controls, which is worse than either
choice made cleanly.
The claim being made is about what the work is, not about permitted use, so it now says that:
developed as academic research, published at ECCV 2026, not the model behind any product.
Anchoring "academic" to the publication makes it checkable rather than asserted.
docs: state the scope of the release as its own section
The research-only point was a clause inside the training-data note, where it read as a caveat
about the data rather than a statement about what this model is. It applies to the whole
release, so it gets its own heading above.
Phrased as three plain facts. A model card is where a reader expects to find scope and licence
statements, so this is the right place for it to be explicit; the arXiv comments field, by
contrast, gets one line and no defence.
docs: disclose the training data composition and SportsShot's licence
Lists the four public shot-boundary datasets used in training, their share of the sampling
distribution, and the fact that SportsShot is CC BY-NC 4.0. The attribution is owed under that
licence's BY term regardless of how the derivative-work question over model weights resolves, so
naming it costs nothing and settles the cheap half of the obligation.
Also states plainly that this checkpoint is a research artifact and not the production model,
which is true and was previously implied to be the opposite by a line on the project page.
Kept as fine print rather than a headline section: it is licence context, not something a user
needs in order to run the model.
Bring back the dark-mode logo inside the flex row
Dropping <picture> was collateral from getting the logo onto the title's line, and it cost the
light/dark switch here. The element itself was never the problem: the heading was, and the row is a
plain div now, so picture goes back in as the flex item with the image sized inside it.
If this renderer strips <source>, the black logo will show in dark mode too and the fallback is a
single logo legible on both backgrounds.
Lay the title and logo out with flexbox instead of a heading
The logo ended up on the badge row, which means this renderer closed the h1 before the image and
left it as a following sibling for the float to catch. Collapsing the markup to one line did not
change that, so the heading itself is the constraint: it will not keep an image inside it.
A flex container sidesteps it. The row is guaranteed, space-between puts the title left and the
logo right, and align-items centres them vertically, which the float never did. The title is a
styled span rather than an h1 for the same reason the image could not stay in one; the model page
already shows the repository name as its own title.
Inline style only, which this renderer preserves. GitHub is untouched and still uses align and
height attributes.
Collapse the heading to one line so the logo stays on it
The logo kept landing on its own line here while the same markup was fine on GitHub. The likely
cause is the heading spanning several indented lines: this renderer breaks multi-line raw HTML
inside a heading into separate blocks, so the image left the h1 before any CSS applied.
Everything is on one line now, and the anchor and picture wrappers are gone since each was another
element the parser could unwrap. That costs the click-through to heygen.com/research and the
dark-mode variant here; GitHub keeps both.
Layout comes from inline style, which this renderer preserves and GitHub strips, so the two now
differ on purpose.
The logo sat on its own line because this card's prose CSS lays images out as blocks, and it drops
the align and height attributes that GitHub honours. Inline style survives here, as Qwen's own card
shows, so display, float and height are set that way instead. The bitmap goes back to 2x now that
its height is pinned in CSS rather than by its own pixel dimensions.
Put the logo back on the right, and match its pixels to its display size
This card drops the height attribute, so the 3x bitmap rendered at its intrinsic 132px, four times
the height of the title text. The file is now 75x44, the size it is meant to appear at, so it is
correct with or without the attribute.
Same header as the GitHub repositories. Without the float the logo joins the heading's line box, so
the rule falls under both instead of through the logo.
Rasterise the logo and size it to clear the heading rule
The SVG logo did not display on this card while the same markup was fine on GitHub. Two things
could cause that, the image format and the relative path, and the page is not publicly reachable
to tell them apart, so both change: PNG at 3x, and an absolute resolve URL instead of a repo
relative one.
Height is 44px, matching the GitHub repositories, which is the tallest the logo can be and still
sit above the rule the heading draws.
Put the title and logo on one line, logo at half size
Same header as the GitHub repositories. The logo was 240px wide on a 241x143 source, so it stood
nearly 142px tall and pushed the six-channel warning down the card. It is now 120px, floated right
inside the heading.
Same header as the GitHub repositories, so the three do not drift: title first and left-aligned,
badges under it, subtitle after the badges. The affiliation line comes out; it is on the project
page, and dropping it keeps the six-channel warning nearer the top where it does its job.
Front matter is untouched.
The card was three lines of front matter. It now describes what the checkpoint is, and leads with
the thing a reader has to know before anything else: the vision tower takes six channels, RGB
concatenated with an optical-flow visualisation, so a standard AutoProcessor pipeline feeds three
and returns nonsense. The weights load either way, which is what makes that failure worth a
warning rather than a footnote.
No standalone transformers snippet is given, on purpose. A three-channel one would run. Instead the
usage section is the real path end to end, including the cuDNN pin that a uv sync silently undoes
and that a Conv3d patch embed cannot do without.
Metrics are not quoted here. Comparisons live on the project page with the protocol attached, and
the evaluation README states plainly which released artifacts are missing for a bit-exact
reproduction of the published figures.
