Stop Banding After You Grade AI Skies
Generated skies posterize under contrast. Use a 10-bit intermediate, dither, and a stills-plus-motion checklist before you add grain or reroll.
Every guide, comparison and workflow we’ve published on Post Production.
33 articles — page 1 of 2
Generated skies posterize under contrast. Use a 10-bit intermediate, dither, and a stills-plus-motion checklist before you add grain or reroll.
Ship masters with an xxHash manifest and packing list, plus a rule for WeTransfer versus Aspera, Signiant, MASV, or a physical drive.
Specify closed GOP, a one-second I-frame interval, and a transcode-on-ingest so a human editor can recut generated MP4s without a render farm.
Relink rules, hashes, Rec.709 tags, and a missing-media protocol for when a regen changes the filename the online expects.
A 30-second false-color pass that flags clipped highlights and muddy faces before any LUT, with the IRE bands to watch on generated skin and chrome.
Prompted grain is baked and fights the compressor. A post stack for size, amount, and chroma you can turn down per platform, after legalizer and loudness.
Resample every 44.1 kHz generate on ingest, use a real anti-alias filter, and keep the project at 48 kHz / 24-bit so export does not click.
Generated clips arrive display-referred. Read Rec.709 waveform, RGB parade and vectorscope, then grade exposure, balance, sat, look.
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.
Park on the splice and match black point, grain size, sharpening, and cadence so a generated insert stops announcing itself next to live action.
Netflix still wants an IMF package, not an MP4. Walk the CPL, PKL, markers, and supplemental versioning, and what a small studio can build without an IMF house.
EBU R 128 is -23.0 LUFS. Netflix near-field is -27 LKFS, dialog-gated. Apple Podcasts is about -16 LKFS. YouTube publishes no integrated LUFS target at all.
Platform MP4s are not masters. Compare ProRes 422 HQ, DNxHR HQX, and FFV1 for 10-bit 4:2:2, alpha, audio tracks, and a file you can still open in five years.
Fold 5.1 and Atmos beds with centre-channel priority, discard the LFE, and check the result on a phone speaker so the words still survive.
Catch inter-sample peaks with a true-peak limiter at −1.0 dBTP before AAC, because a brickwall at 0 dBFS still fails QC after the encode.
Video models render at 24fps and Versely timelines default to 25. The three ways to conform, the arithmetic behind each, and the one that wrecks lip sync.
Background text a model got wrong is not repairable, but it is not necessary either. A soft-focus pass that reads as optics, and where it stops working.
An intake check, an input-prep chain and a three-way decision for when voice conversion beats re-recording or generating the line from text.
A short dissolve at the join is the standard fix for exposure and identity drift between chained clips. Dissolve lengths, and when a hard cut still wins.
Generated footage reads wrong because imperfection is absent, not detail. A finishing order for grain, vignette and flare, and the amounts that hold up.
Rapid consonants expose a 1–3 frame lag slower speech hides. Target a 120ms window, nudge the clip, and rewrite lines that keep tripping it.
Most lip-sync drift is an export fault, not a model fault. Re-export at native frame rate, skip optimized rendering, and feed uncompressed audio.
A match cut hides the join between two generations by carrying a shape or a motion across it. How to plan the pair in the prompt and which frames to trim.
Mixed edits break on grain, grade, focal length and motion blur. The order to fix them in, and the one thing you cannot repair in post.