Versely
    AI Models

    Luma Ray 3: HDR when audio is someone else's job

    16-bit HDR 1080p, cheap entry, no native audio. Grade-friendly plates you will score later.

    Versely Team••6 min read

    Do not pick Ray 3 for a talking ad. Pick it when the colour volume is the deliverable.

    Luma Ray 3 generates 16-bit HDR at 1080p, at a cheap entry, with no native audio. The file is a grade-friendly plate. Someone else — you, in the session, with stems — will score it. If the brief is a mouth saying a line, this is the wrong row. If the brief is a master that still has highlight texture after the SDR downmap, this is the row.

    What "HDR plate" means in practice

    Ray 3 is the rare generate that is worth thinking of as negative, in the photochemical sense: a camera original with headroom, not a finished Rec.709 YouTube file. Luma's own pitch is native 16-bit HDR, 1080p, EXR-class colour for a grade. That is a finishing claim, not a social-clip claim.

    Most of the catalog still emits an SDR-looking clip. You can tag HDR metadata onto that clip. You should not. Stretching Rec.709 onto a PQ curve does not invent speculars. Grade HDR once, keep the SDR downmap is the finishing doctrine; Ray 3 is the generate that doctrine can actually use, because the highlight was in the render.

    Cheap entry matters here. HDR is a reason people over-generate on hero-priced models and then discover the talking track is unusable or the identity drifted. Ray 3 is allowed to be the mood, the sky, the product chrome, the wide you will sit under a real mix. It is not allowed to be the presenter.

    No native audio is the other half of the honesty. There is nothing to lip-sync. There is no 48 kHz dialogue hiding in a toggle. Plan the soundtrack as a separate job: VO, licensed bed, Firefly-licensed bed if the brief is commercially safe wallpaper, a real mix if the brief is a film. The plate does not owe you a voice.

    When the colour volume is the brief

    You are in Ray 3's lane when someone has asked for HDR by name, or when the shot's value is what happens to light:

    • Chrome, windows, practicals next to a face you will not generate as a talker.
    • A sky that has to survive YouTube's SDR downmap with a shoulder, not a slab.
    • B-roll and establishing shots where a bed will be laid in the edit.
    • Mood-board passes you expect to grade, not to post raw.

    You are out of the lane when:

    • A line of copy has to be said on camera.
    • The product label is contractual and the still already exists — that is an I2V lock, usually Seedance.
    • You need 4K as the camera original. Ray 3's native generate is 1080p HDR; 4K is a later upscale conversation, not the reason to pick it.
    • The deadline is "export and upload" with no grade.

    "The client has an HDR TV" is not a brief. "Deliver HDR10 plus an SDR 1080p" is a brief. Answer the second with a 16-bit plate and a downmap QC. Do not answer it with Veo audio-on 4K and a hope.

    Audio is someone else's job — on purpose

    The mistake is treating missing audio as a defect to patch inside the generate. Prompting Ray 3 to "add a voiceover feel" does not produce speech. It produces a silent clip whose motion you then have to fight with a dub.

    Do this instead:

    1. Generate the plate on Ray 3. Judge it on a waveform and a pair of stills, not on whether it "feels like an ad" with the speakers muted.
    2. Build the soundtrack in the editor: VO, bed, SFX timed to the picture you already like.
    3. If a later shot needs a mouth, that shot is Veo or Happy Horse. Cut it against the Ray 3 wides. Do not force the wide to speak.

    This is ordinary picture-then-sound production. Native-audio models inverted it for talking clips, which is correct for talking clips. Wides never needed the inversion.

    For generation when you do not yet care which specialist owns the file, the door is the AI video generator. Open Ray 3 after the brief has named HDR or named a silent plate.

    What not to steal from the arena

    Ray 3 will not win a T2V-with-audio board. It cannot. The board requires a soundtrack. A control or colour model losing that vote is not information.

    It will also lose a talking-head bake-off, a 30-second identity bake-off, and a six-shot gag bake-off. None of those are the colour-volume job. Keep the specialist in the specialist's lane and the Elo screenshot off the routing doc.

    FAQ

    Can I use Ray 3 for a talking product ad if I add TTS later?

    Only if we do not see the mouth, or you are prepared to treat the talking beat as a different generate. TTS on a silent wide is fine. TTS on a generated face is a lipsync problem Ray 3 did not sign up for. Generate the talker on a native-audio model.

    Is 16-bit HDR the same as uploading a "HDR" tag?

    No. Ray 3's claim is native 16-bit colour in the generate, gradeable, with real highlight texture. Tagging an SDR file as PQ is a metadata lie. Grade Rec.2020 PQ or HLG from a plate that had the headroom.

    Why not just generate HDR on Veo 3.1 at 4K?

    Veo is the dialogue and 4K row. Its job is the mouth and the master when speech cannot be optional. It is the wrong reason to pick a colour-volume plate, and a talking generate is a poor HDR original. Split the jobs: Ray 3 for the gradeable wide, Veo for the line.

    Does Versely list a /models/luma-ray-3 page?

    Not as a head slug in the current catalog. The routing rule does not depend on that URL. When you need the HDR silent plate, pick Ray 3 by job. When you need a generic generate door, use the video generator and then name the job.

    Colour volume is a finishing brief. Speech is a different brief. Ray 3 only answers the first, cheaply, at 16-bit 1080p, with no soundtrack to forgive a wrong pick.