How to show color grading work online without compression
Send one file, untouched, and prove what arrived.

The reason is mechanical rather than rhetorical. A rate-limited encoder allocates bits according to what is in each frame. An ungraded log shot is flat, low in contrast and cheap to encode. A graded shot with crushed blacks, a bloomed highlight and grain running through the whole frame is expensive. Feed both to the same ladder under the same ceiling and the expensive half is the half that gets approximated away. The delta the viewer sees is your work plus an encoder’s opinion, and there is no way to separate the two after the fact.

So the first decision is the host, and it is a binary one: either the platform re-encodes your upload into renditions, or it stores your file and serves those bytes. Everything else on this page, the way you cut the comparison, the codec you finish to, the checks you run, only matters once that step is off the table.

Shadow gradation. A lifted black, a slow roll into a night exterior, a face held two stops down. The information there is small differences between adjacent dark values, and coarse quantization is the act of throwing away small differences. The result is banding across the falloff you spent the longest shaping, and it shows up on the client’s screen and not on yours.

Grain. Grain is high-frequency noise across the entire frame, which defeats the frame-to-frame prediction every modern codec depends on. An encoder under a ceiling has two options: spend real bits everywhere at once, or smooth it. It smooths it. A film-grain emulation you tuned to a specific texture arrives as a soft wash, and the client tells you the look feels digital.

Halation. The bloom around a hot highlight is a wide, smooth, low-contrast gradient with a colored edge, sitting next to the brightest thing in the frame. That is the same expensive case as a sky, made harder by the neighbouring contrast. It is also the part of a look a client notices without having a word for it, so when it flattens they report the image as ordinary rather than as compressed.

Two numbers from that specification are worth carrying into a delivery decision. The documented maximum resolution for that decoder is 4096 by 2304 pixels, which fits UHD at 3840 by 2160 and DCI 4K at 4096 by 2160 and nothing larger. And hardware acceleration is guaranteed only to 1920 by 1088; above that the documentation says decoding uses DXVA if the hardware supports it and software otherwise. A client reporting that your 4K reel stutters while the fan runs is usually describing a software decode, not a bad file.

On 10-bit 4:2:2 specifically, the discipline is to say what is written down. Nobody documents 10-bit 4:2:2 HEVC decode in a video element, in either direction. Apple documents HEVC 10-bit 4:2:2 as an encoding capability added in Compressor 4.6.1 on Apple silicon Macs running macOS Monterey 12.3 or later, which is a statement about producing the format and not about a browser reading it. The full reading of every vendor document is in HEVC 10-bit 4:2:2 browser playback.

For showing the work rather than sending it to one client, the portfolio page is built from the films already in your library. Tag one Hero and it runs full width, tag two Featured and they sit side by side, everything else lands three across, in the order you choose. It lives at your own address, yourname.uncompressed.io, and you can drop text, images, buttons and a contact form between the rows. Autoplay is off by default; when you turn it on it applies only to Hero cards, muted, one at a time. The longer version of that setup is in video portfolio hosting.

The cost of this shape of host, stated plainly, because it is real. With no ladder there is no lower rung, so a viewer on a thin connection buffers instead of quietly sliding to a softer version of your grade. For a public showreel aimed at an unknown audience on unknown devices, an adaptive platform is the better engineering and you should use one. For a before and after going to people whose judgement of the picture is the point, buffering for ten seconds and then seeing the actual image is the better failure mode.

Removing the re-encode does not remove the viewing environment, so choose the environment before you finish the export. A reel that will be opened on laptops in offices is watched in a bright room, on a panel that is probably brighter and cooler than your reference, by someone who will not dim the lights. A pass that reads beautifully at 100 nits in a dark suite can read as murky there, and the client will describe your careful shadow work as underexposed.

The practical answer is not to compromise the grade. It is to make the version you send match the room it is going to. Keep the reference pass for the deliverable and the archive, and build the web reel with a slightly lifted floor if the audience is unknown, checking it on a normal consumer laptop before it goes out. Say which one the viewer has, in one line next to the link. A colorist who states the intent removes an argument that would otherwise arrive as a note.

The same logic applies to the split screen. If the ungraded side is log, tell the viewer it is log rather than letting them read it as a failed grade. If it has a transform applied, say which. A comparison with a stated method is evidence. A comparison without one is a picture of two videos.

Read together, the whole method is short. Build the comparison as one file so the halves are treated identically. Finish to a codec the receiving machine documents, and keep 4:2:2 for the download rather than the review link. Host it where nothing re-encodes it. Verify the delivered file rather than the marketing. Then reserve the calibrated room for the one session that decides the show. The neighbouring answer page, video hosting for colorists, covers the hosting side of that in more depth, including the published delivery ceilings on the platforms that do re-encode.