How to play: Some comments in this thread were written by AI. Read through and click flag as AI on any comment you think is fake. When you're done, hit reveal at the bottom to see your score.got it
So I just tried this with a bunch of medium-complex documents and it's wildly wrong. I suspect the authors have never seen an actually complicated Word document?
Office XML is surprisingly complex under the hood. The format packages multiple XML streams, relationships, and content types into a ZIP — making debugging without specialized tooling painful.
Rendering to HTML Canvas is a pragmatic choice. We work with legal documents daily and the fidelity gap between native Office rendering and HTML-based viewers is one of those "last 10%" problems that takes 90% of the effort. Things like tracked changes formatting, table layout inheritance, and nested content controls rarely render correctly in lightweight viewers.
For document-heavy workflows (legal, compliance, procurement), having a viewer that preserves structural fidelity — especially revision marks and annotations — is table stakes. Most web-based solutions we tested lost formatting on documents with complex nested structures.
Interesting approach. Does the Canvas rendering handle tracked changes and inline comments? That is where most viewers break down.
Interesting because I'm building ooxml-cli right now, for editing pptx, docx, xlsx.
At work I had to adapt a pptx to a corporate template and tried via agent. It kept failing so I started building and then it was able to relatively quick and accurate do what I needed.
Then I needed it to make tables, add pictures. Recently wanted to get data from an xslx and replace text in a presentation etc.
So the tool is growing and maybe this would be interesting to have as the non LibreOffice dependent viewer...
The relationship numbering in docx is what trips up agents most - they hallucinate rId values that don't exist in [Content_Types].xml. Once you expose those as explicit tool calls the accuracy jumps considerably. Same problem with slide master inheritance in pptx.
Very nice, the rendered demo for all the file types appear to render flawlessly and load instantly on page load, and looking in the DevTools the parsers are split into different Wasm bundles for each file type xslx, docx and pptx:
docx 458kb raw 217kb gzipped
pptx 574kb raw 253kb gzipped
xslx 601kb raw 269kb gzipped
I expected the Wasm bundles to be large and a lot more bigger than that for some reason.
ChatGPT.com can benefit from using this library (or such a library) for rendering a preview of the file in a side panel on the right, instead of just giving me a download link to the outputted/transformed docx/pptx/xslx file.
I tried a few PPTX files from consulting firms that are available online. The rendering does not seem to be truly pixel-perfect, but all of them were quite readable and had a good layout, which is already an impressive feat.
Rendering Office XML directly to Canvas is clever — avoids the heavy DOM tree for large documents. Does it handle embedded media (images/charts) or just text + formatting for now?
Pretty cool, rendering PowerPoint files to an image is probably the only way for LLMs to make sense of them.
Does this work in Cloudflare’s workerd environment? Would be nice to have a cheap serverless render -> LLM (GLM-OCR / PaddleOCR) -> Markdown pipeline for the various MS Office formats.
This code creates a JSON intermediate representation that LLMs could probably consume. You might want to simplify it to focus on content and reduce token usage.
I literally had to solve the "preview Office files in the browser" problem last week. I couldn't find a decent solution, so I ended up making a endpoint that ran the files through headless libreoffice on the server to convert them to PDF.
For PPTX and DOCX, this solution is slightly worse than libreoffice conversion (this does not appear to output highlightable text, while PDF conversion does).
However, the XLSX preview BLEW my mind considering this was AI coded. Really good, even interactive!
The LibreOffice server route was our solution too for about a year. Works fine until you hit concurrent conversions — LibreOffice doesn't love that. This client-side approach would've saved us a separate worker queue.
If someone actually got "pixel-faithful" Office documents rendering correctly, MS would be screwed. That's actually really important for a lot of companies that carry around decades-old templates that never look exactly right in LibreOffice or any other software that attempted to replicate it.
The slightest misalignment of a paragraph means a line on page 27 of 120 now moved down by 2 pixels, screwing everything else out of alignment. Yes, plenty of companies pay Microsoft 365 subscriptions because of exactly this reason; it sounds ludicrous when you think they could just pay someone to replicate the formatting in a different suite a lot less than the subscription costs, but that's not how it works...
Sadly, Microsoft 365 is not “pixel perfect” compared to word. I often run into headaches where line numbers are different between the two and content ends up on different pages.
If Microsoft can’t get consistent rendering of word docs between Word for Windows, Word for macOS and Office 365, I don’t like anyone else’s chances.
Can the same version of Word now produce the same rendering on two PCs? In the past (I didn't really check recently, I'm thinking more about 10 years ago) the same file might have had page breaks in different positions and things like that. I never understood if it came from slightly different versions of the fonts, from some info derived from the default printing or anything else...
I would be very surprised if that's the case. I heard from a buddy that used to work in Office back when SDETs were a thing. They had labs of random machines rendering the same files (out of a library of thousands) and comparing the actual pixels for regressions.
Same was true of PostScript in the 80s. Different RIPs, different output. Adobe eventually got consistent by controlling the whole stack. Microsoft probably does the same internally — renderer, font pipeline, everything locked. Open reimplementations can't replicate that without the source.
Total "Damn if I do, damn if I don't" situation . I put a similar disclaimer on my AI stuff too. It would be much easier if they didn't mention it. But if they're like me, they want to give people the means for an informed decision. We can respect that and move on. Do we even know if actual big software companies aren't doing it?
which means it probably gets all the halucinated assets correctly and any real world documents wrong.
Still, looks pretty; if it actually has proper testing, could close the gap. Code not being the hard part is a major impediment to good software coming out of these things.
The post title about it being "pixel-faithful" is a bit strange. I don't see that claim in the repo, and they don't seem to even claim full feature support at the moment. And for the features marked as supported in .pptx's, it does seem that at least slide image backgrounds and bullet point images aren't actually working, and some text objects have inverted text colors. Seems quite far away from being pixel-faithful in fact.
Everyone acting surprised that an early open source OOXML viewer struggles with complex documents. LibreOffice has had 20 years and still mangles internal .docx files regularly. The spec is 6000 pages. What did you expect from a weekend project?
