Versely
    Guides

    Grade HDR Once, Keep the SDR Downmap

    Grade Rec.2020 PQ or HLG once, roll off highlights, and QC two stills of the same frame so the SDR YouTube downmap does not crush the HDR master.

    Versely Team••9 min read

    Most AI pipelines export SDR. The picture is a finished Rec.709-looking frame, already compressed, and that is what you should deliver unless someone has asked for HDR by name. When they have, do not grade an SDR file, slap HDR metadata on it, and hope. YouTube's HDR upload guide is blunt: grade in Rec.2020 with PQ or HLG. Using a different configuration, including DCI P3, produces incorrect results. YouTube will also build an SDR downconversion for everyone who is not on an HDR screen. If you never look at that downmap, the version most people watch is a surprise.

    The job is one HDR grade, a highlight roll-off that still has shape after the downmap, and two QC stills of the same frame: the HDR master and the SDR it becomes.

    Grade in Rec.2020, deliver two masters

    ITU-R BT.2100 is the HDR-TV production recommendation. It specifies two transfer functions: PQ (Perceptual Quantizer, SMPTE ST 2084) and HLG (Hybrid Log-Gamma). Colour primaries are Rec.2020. YouTube's upload requirements match that: Rec.2020 primaries, Rec.2020 non-constant-luminance matrix, PQ or HLG, 10-bit or 12-bit, at 720p and above. HDR is not supported on YouTube at 480p or 360p, which is one more reason the editor's 480p preview cannot be the HDR QC. That preview is a free pass with a short per-user cooldown; it is for cut and timing. Judge HDR on a 10-bit master.

    YouTube detects HDR metadata and, in its own words, produces "HDR transcodes for HDR devices and an SDR downconversion for other devices." Viewers on non-HDR devices see SDR. That is not optional. The downmap is part of the delivery whether you author it or YouTube does.

    Do not start from a generated SDR clip, stretch it, and retag it as PQ. Stretching Rec.709 into ST 2084 does not create highlight texture. It maps a 100-nit-ish picture onto a curve that can encode up to 10,000 cd/m² and leaves a thin, noisy image that YouTube will then downconvert again. If the client wants HDR, generate the plates as usual, then finish in an HDR project (DaVinci Resolve, Premiere, Final Cut, and After Effects are the authoring tools YouTube lists). Most text-to-video and image-to-video output is still an SDR-looking file. The HDR pass is a grade, not a model switch.

    Keep the SDR YouTube file in mind from the first trim. YouTube's recommended upload encoding wants BT.709 for SDR, with 8 Mbps at 1080p for 24/25/30 fps and 35-45 Mbps at 4K for those same rates. HDR recommendations sit higher: 10 Mbps at 1080p and 44-56 Mbps at 4K for 24/25/30. Versely timelines default to 25 fps, which is in the standard-frame-rate column. AI video resolutions is the generation-side choice; do not jump to 4K only because the master is HDR. YouTube accepts HDR at 720p and 1080p.

    PQ versus HLG

    Pick one transfer function for the project and do not mix them in a sequence.

    PQ (ST 2084) is absolute. Code values map to cd/m², from 0 up to 10,000. A 1,000-nit specular is a 1,000-nit specular on any PQ display that can show it. That is why streaming HDR10 is PQ, and why PQ files need mastering-display metadata. YouTube wants SMPTE ST 2086 for the mastering display, plus CEA 861-3 MaxFALL and MaxCLL. If those brightness numbers are missing, YouTube fills in the values for a Sony BVM-X300. Do not skip the metadata and hope. An untagged PQ file is how you get a picture that looks right on your monitor and wrong everywhere else.

    HLG is relative. It was designed so an SDR display can show a reasonable image without a separate downmap, which is why it shows up in live and broadcast. It is not "free SDR." The SDR-compatible HLG look is not the same picture as a graded Rec.709 downmap, and YouTube will still produce its own SDR conversion. Grade HLG as HDR, then check the SDR you will actually ship.

    PQ (ST 2084) HLG
    Mapping Absolute nits Relative, scene-referred
    YouTube Supported, with ST 2086 + MaxFALL/MaxCLL Supported
    Metadata Required for PQ signalling Transfer + Rec.2020 tags
    Typical use VOD, HDR10 Live, some broadcast
    SDR story You author a downmap, or YouTube does Somewhat viewable on SDR, still check a real Rec.709 file

    YouTube's common-issues note is the other hard rule: do not master in DCI P3 and tag Rec.2020. That pairing oversaturates and shifts hue. Grade Rec.2020. If a monitor is P3-native, that is a display management problem, not a tagging trick. Why colours shift between screen and export is the SDR/P3 version of the same mistake.

    Highlight roll-off that survives a downmap

    HDR headroom is where this grade usually dies on YouTube. On the HDR monitor, a chrome edge at several hundred nits looks expensive. YouTube's automated SDR downconversion then maps that region toward Rec.709 white, and the shape disappears: a specular becomes a white slab, a window becomes a rectangle, a face near a practical goes chalky. YouTube says the automated downconversion "can deliver good results with no effort" and, on challenging clips, "might not deliver the perfect result." Product chrome and blown practicals are challenging clips.

    Roll the highlights in the HDR grade so they still have a shoulder after the downmap. That means:

    • Put speculars on a curve, not on a clip. A small peak that falls off will survive. A plateau will become a slab.
    • Keep faces out of the HDR highlight region. Generated skin already sits smooth. HDR just makes the plastic brighter.
    • Do not park the whole frame at a high average luminance to make the HDR "feel" HDR. MaxFALL exists because a frame that is bright everywhere downmaps to a grey soup or a clip-fest. Let the picture be mostly SDR-bright, with a few true highlights.

    If you want more control than YouTube's automatic conversion, YouTube documents a 3D LUT hint: grade a Rec.709 / gamma 2.4 version of the conversion in the colour tool, export a .cube, and pack it with the HDR file via their metadata tool. The limitation is also documented: no spatial or temporal control, so power windows and per-shot keys will not survive as hints. That is another argument for getting the roll-off into the HDR grade itself.

    What a grade can and cannot rescue in AI footage still applies. A clipped SDR generation does not become HDR by tagging. If the waveform was a hard ceiling in Rec.709, there is no highlight to expand. Regenerate the plate with a dimmer key, then grade HDR.

    Two-master QC stills

    Leave the session with two stills of the same frame, named so a reviewer cannot mix them up: shot04_hdr.png and shot04_sdr.png. Pick a frame that contains the hardest highlight in the scene (chrome, a window, a practical next to a face), not a mid-tone wide that will pass anything.

    QC both stills at 100% on a display you trust, then on a phone. The HDR still is the intent. The SDR still is the version most YouTube viewers will get. Fail the pair if any of these is true:

    1. Speculars that had shape in HDR are a flat white in SDR.
    2. The face that sat in a healthy mid-tone in HDR is chalk or mud in SDR.
    3. A brand colour (a pack, a logo, a wall) rotates hue in the downmap.
    4. Banding that was hidden in the HDR highlight shows up in the SDR sky or backdrop.

    If the pair fails, fix the HDR roll-off and re-export both. Do not "fix" the SDR with a second creative grade that the HDR master does not share. The whole point of grading HDR once is that the SDR is a downmap of that grade, not a fork.

    Export paths: HDR as 10-bit HEVC, VP9 Profile 2, or AV1, tagged Rec.2020 + PQ or HLG, with ST 2086 and MaxFALL/MaxCLL if PQ. SDR as Rec.709, legal range, at the bitrate YouTube recommends for that resolution. Keep the two files together in the delivery folder. The video editor is where the cut is built; the HDR grade and the downmap stills live in the finishing tool you actually have scopes on.

    A generated clip that never needed HDR should stay SDR. "The client has an HDR TV" is not a brief. "Deliver HDR10 plus an SDR 1080p" is a brief. Answer it with two masters of one grade, not with metadata on a stretched Rec.709 file.

    FAQ

    Can I upload the HDR file and let YouTube make the SDR?

    Yes. That is the default. You should still export your own SDR downmap still (and, for anything with chrome or windows, a full SDR file) because YouTube's conversion is automatic and, on hard highlights, not guaranteed to match what you signed off on the HDR monitor. The still is how you see the miss before the client does.

    Is HLG "safer" because SDR TVs can show it?

    Safer for live. Not a substitute for a Rec.709 file. HLG's SDR-compatible rendering is not your graded SDR master, and YouTube will still downconvert. If the brief includes a YouTube SDR audience, check a real Rec.709 downmap.

    Do generated clips contain HDR metadata I can keep?

    Treat them as SDR-looking files unless a probe of that exact export says otherwise. Do not assume a model's marketing page. Probe the file, then build the HDR grade in Rec.2020 PQ or HLG in the finishing tool.

    Why not grade SDR and HDR as two separate looks?

    Because they will drift. Two looks is how the YouTube SDR file stops being the HDR film. Grade HDR once, downmap, and fail the pair when the downmap crushes what the master showed.