Novus Examples grows up: 401 video fixtures, 480 templates, and OBS-compatible scenes

Novus Examples has had a large week. The catalog is now 2,428 files across seventeen categories and 238 documented formats, up from 2,097. The visual template library went from 180 templates to 480. The in-browser editor learned to open and rebuild archives instead of showing you a hex dump. And the four broadcast template categories now ship something new alongside their artwork: an OBS-compatible scene collection, so an overlay you download arrives as a scene rather than as a PNG you still have to place.

None of that is interesting as a set of numbers, so this post is about the reasoning instead. Why video needed rebuilding rather than topping up, why 300 new templates required new drawing capabilities before a single one could be generated, and why a scene collection has to ask you to relink a file — which looks like a defect and is in fact the only honest option available.

Video had 70 files. On paper that is a category; in practice it was the weakest part of the library, because it had the most subcategories of anything in the catalog and the fewest files in each of them. A subcategory with three files in it does not let you isolate a variable. It lets you confirm that a format exists.

The deeper problem was that the tasks people most want to test had no coverage at all. If you are working on video upscaling, denoising, frame interpolation, matting, or depth estimation, what you need is not a clip. What you need is a degraded clip and the pristine clip it was degraded from, so that when your model produces an output you have something to compare it against. Without the reference, "did that work?" is a matter of opinion, and opinions do not catch regressions.

So the expansion was built around pairing. Ten AI-video task suites now ship where every degraded input records the identifier of its own ground truth, and the two share a group so you can navigate from either side. The encodes are deliberately two-tier — ground truths are encoded at a higher quality than the inputs — so the measurement floor sits below the effect you are trying to measure, and compression artefacts in the reference do not become the limit of what you can detect.

The rest of the programme fills in what a video pipeline meets in production: fourteen codecs, seven pixel formats, playable HLS and DASH packages, captions in the encodings that actually break parsers, multi-track accessibility fixtures whose dispositions are genuinely set, chapters in six formats, and a set of intentionally-corrupt containers. Video is now 401 files and the second-largest category in the library.

The detail worth stealing from this work, if you build fixtures of your own, is the container set. Seven containers hold the same footage, and they hold it because one encode was stream-copied into each of them rather than encoded seven times. The video bitstream inside every one is byte-identical.

That sounds pedantic until a test fails. If each container had been encoded separately, a failure on the MKV could be the container, or it could be that the MKV encode happened to use a different GOP structure. You would have to investigate to find out. With a stream copy, the container is the only variable that exists, and the result interprets itself.

The same discipline runs through the rest. A denoise fixture states its noise level rather than being "noisy". An interpolation fixture states which frames were removed. The point of a fixture is not to be realistic — it is to be known.

The visual template library had 180 templates and one composer behind them: a background, a pedestal, a product, a caption, a logo. That is a perfectly good picture, and it was the right picture for the categories the library started with, which really were "a product on a background".

Running 300 new categories through the same composer would have produced 300 more of the same picture with different titles. That is volume rather than coverage, and because template slugs are permanent, it would have been a permanent mistake. So the capability gaps were closed before any of the new rows were generated, and the result is seventeen genuinely different layout families rather than one repeated: transparent broadcast overlays, alert cards, avatar state rings, table sheets, flow diagrams with real arrowheads, certificates, badges, signage, shelf labels, display ads, quote cards, and hero splits.

A few of those gaps are worth naming because they were invisible failures rather than missing features. Diagram templates had been faking their connectors with thin rectangles, because the schema had rectangles and ellipses and nothing else. A circular avatar was impossible, which quietly ruled out about a third of the new categories. And every one of the original 180 scenes named a system font that the export could never actually load — an SVG rasterised through an image element runs in a restricted mode and refuses to fetch fonts — so they exported in whatever face the rendering context happened to have. Fonts are now embedded, and the same bytes draw both the committed preview and the Studio canvas.

The quality gates are code rather than eyes, because at 300 families spot-checking is not a control. Each family declares an aspect band and fails loudly if a row is composed for the wrong shape. The composer asserts that no two text, image, or table nodes overlap — overprinted type depends on how long the words happen to be, so it survives every spot-check and then ships on the row nobody looked at. And a sweep across all 300 generated scenes found 50 truncated strings, things like "Book at example.c…", which are now zero.

A template library that can produce a certificate can also produce a fake credential, and the honest response to that is a constraint rather than a disclaimer nobody reads. All fourteen certificate and badge scenes carry a SAMPLE marker that is locked — it cannot be edited or deleted in the Studio — and a fictitious issuer. The legal, financial, healthcare, fitness, and non-profit categories all carry a locked disclaimer. Barcodes encode nothing and use a restricted-circulation prefix. Healthcare templates are operational only: there is no intake form and no medical-history form anywhere in the set.

A brand-safety pass over all 300 rows also turned up three genuine trade-dress leaks — safe-area dimensions that were not generic facts but a specific platform's exact spec. Dimensions in general are facts and are fine; those three were the template, and they are gone.

One bug from that audit is a good reminder of how these things fail. The badge family had the SAMPLE marker and the subtitle as the same node, which meant a row supplying its own issuer copy silently overwrote its own marking. It looked fine in every preview. They are two nodes now, and the marker is locked.

The complaint that started this one was that the editor could not produce all the file types it displayed. That was true, and the real number was worse than the complaint: nineteen of 216 formats could be created from nothing. The editor opened nearly everything it reasonably could and authored almost nothing.

Archives were the biggest gap, because so many everyday formats are archives wearing a different extension. A .docx is a ZIP. So is an .xlsx, a .pptx, an .epub, an .odt, a .cbz, and a .jar. Open one now and you get a member list rather than a hex dump; open word/document.xml inside it, edit the XML, and the rebuilt file still opens in Word. Dispatch is on the file's magic bytes rather than its extension, so a renamed archive still opens, and the cases that cannot be handled safely — ZIP64, encryption, bzip2, xz — are refused by name and fall back to the hex inspector rather than producing something subtly wrong.

The editor can also author binaries now. A JSON tone spec renders to real 16-bit PCM WAV with a seeded generator, so the same spec always produces byte-identical audio. A self-contained glTF packs to binary GLB with correct chunk padding, and one with an external buffer reference is refused with the offending URI named rather than producing a GLB with a dangling pointer. The New menu went from 13 starters to 51.

Two silent data bugs got fixed along the way, and both are the kind worth watching for in your own exports. The Document Studio was writing every spreadsheet cell as a string, so an invoice row landed as four text cells, Excel showed "number stored as text", and SUM returned zero. Cells are now typed — numbers, currency including accounting negatives, percentages, ISO dates, booleans — and deliberately not typed where conversion would be data loss: leading zeros, sixteen-digit card numbers, and phone numbers with a leading plus stay as text. And the Forms Studio's JSON export could not round-trip even in principle, because the export stripped the internal ids that the import guard required. It has an Import button now, and the ids are minted on the way in.

The last thing to land is the one most likely to save someone real time. The four broadcast template categories — live stream overlays, stream alerts and motion frames, stream scene backgrounds, and avatar state sets, 48 templates in total — now ship an OBS-compatible scene collection alongside their PNG and WebP downloads. You take it through Scene Collection → Import, and you get a scene with the overlay already placed above a capture placeholder, bounded to the artwork's own canvas so it scales sensibly if your base canvas differs.

It is described as OBS-compatible rather than as an OBS file for a reason: Novus Stream Solutions is not affiliated with or endorsed by the OBS Project, and the file ships no software and needs no plugin. It is plain JSON in a documented format, readable in a text editor.

Now the awkward part. On import, the image will show as missing, and you will have to relink it once. This is not a shortcut we took. OBS stores an image source's path as an absolute local path on your machine, and a file downloaded through a browser has no way to know where it landed — so there is no value that could have been written into that field that would be correct on your computer. Guessing produces a path that is wrong everywhere and fails confusingly.

So the collection does the opposite of guessing. It carries a folder name that is obviously not a real location, which means OBS immediately raises its own Missing Files dialog, and that dialog has exactly the tool needed: "Search Directory…" relinks every source whose filename matches a file in the folder you point it at. One click, whole collection.

That only works if the filenames match, which forced a second change worth mentioning. Every template's primary download lives at a permanent URL ending in canonical.png, so a plain download attribute saved all 480 of them as canonical.png — download five and you had canonical(1) through canonical(4). Downloads now carry a suggested name derived from the template slug and its dimensions, and the scene collections reference exactly that name. Fixing the annoyance and enabling the relink turned out to be the same fix.

The catalog stands at 2,428 files across seventeen categories, with 238 formats marked available on the coverage page because a real documented file exists for each, a short list marked as coming, and about fifteen marked unsupported with the actual reason given rather than silence. Alongside it are 480 visual templates across 40 categories, 42 curated document starters, twelve live test targets, and an editor that now reads archives and writes binaries.

The shape of the thing has not changed, though. It is still a library you open, take a file from, and leave — free, no account, nothing uploaded. The additions are all in service of the same question the library has always been trying to answer quickly: what is the exact file that would prove this works?

For orientation, Novus Examples covers everything in one place and Tool maps is the current map of the app. The two new walkthroughs are Test a video pipeline with the 401-file video set for the video set and Build an OBS scene from a broadcast template for the broadcast templates.