Grade Generated Footage on Rec.709 Scopes
Generated clips arrive display-referred. Read Rec.709 waveform, RGB parade and vectorscope, then grade exposure, balance, sat, look.
A print-emulation LUT is the wrong first move on generated footage. The clip is already a display-referred Rec.709 picture: contrast is in the file, highlights are already rolled, and there is no log encoding waiting to be developed. Drop a film-print cube on that and you stack an S-curve on an S-curve. The mids go muddy, the skin rotates, and you spend the next hour undoing a look you never needed.
Scopes are the first move because they tell you what the file actually contains, not what the thumbnail suggested. Set them to Rec.709 legal range, read three of them, and then grade in a fixed order: exposure, balance, saturation, look. Twenty minutes is enough for a short sequence if you do not start with the look.
Generated footage is already display-referred
Camera log (S-Log3, LogC, V-Log) is scene-referred. A print-emulation LUT is built for that input. It expects a flat image whose contrast will be created by the LUT. Generated video does not arrive that way. Image and video models write a finished-looking frame, usually tagged or treated as Rec.709 / sRGB, already compressed. What a grade can and cannot rescue in AI footage is the limit case of the same fact: the grade can only redistribute values that exist.
That is why a Kodak-2383-style print LUT, or any "cinematic" cube aimed at log, fails as step one. It is also why a generation-time style preset is not a substitute for a grade. A preset is a sampling bias. It does not put two models on the same baseline, and it does not put the file into a working space a print LUT was built for.
Treat every generated clip as Rec.709 display-referred until a probe of the file says otherwise. ITU-R BT.709 is the HDTV studio standard: in 8-bit it maps reference black to code 16 and reference white to code 235; in 10-bit those same points are 64 and 940. If your scopes are set to full range (0-255 / 0-1023) while you grade a legal-range file, the picture will look lifted and you will "correct" it into crushed blacks. Set the scope, then look at the picture.
Assemble the sequence first. The video editor is an EDL-based timeline; preview: true gives a free 480p pass with a short per-user cooldown, and the final export is charged once regardless of clip count. Do not grade the 480p preview. Export the matched sequence, then open that file in the grading tool where the scopes live.
Three scopes on Rec.709 legal range
Open waveform, RGB parade, and vectorscope. Set the working colour space to Rec.709 and the scale to legal (studio) range, not full. If the tool has a separate "video vs data" or "limited vs full" switch, it has to agree with how the file is actually encoded. A mismatch here is the whole defect.
Waveform (luma). This is exposure. On a Rec.709 legal waveform, reference black sits at 0 IRE (code 16 in 8-bit) and reference white at 100 IRE (code 235). You are looking for three things: where the face sits, whether highlights are a peak or a hard ceiling, and whether shadows have a floor. A ceiling at 100 IRE with no texture is clip. A pile-up on the 0 IRE line is crushed black. Generated talking heads often sit too high: the face is bright, even, and plastic, because the sampler favoured a "well-lit" still. Generated product chrome often has a flat white plateau instead of a small specular peak. Both are exposure problems, not look problems.
RGB parade. This is balance. Three waveforms, one per channel. On a true neutral (a white shirt, a grey wall, a product card) the three traces should sit on top of each other at that region. If red rides above green and blue on the shirt, the shot is warm. If blue rides high in the shadows, the shot has a cool cast. Mixed-model timelines fail here first: two families will not share a white point, and a look LUT will not invent one. A creative LUT will not fix mismatched shots is the matching order; the parade is how you see the mismatch.
Vectorscope. This is hue and saturation. Skin should sit on the flesh line (the I-line, around eleven o'clock on a Rec.709 vectorscope). Saturation is the distance from centre. Generated clips often arrive oversaturated, especially at high CFG, so the blob smears toward the box. Pull sat until skin holds the line and the rest of the frame is not neon. Do not use the vectorscope to apply a teal-and-orange look. That is step four.
| Scope | Set to | Pass if | Fail if |
|---|---|---|---|
| Waveform | Rec.709 legal | Face in a consistent luma band; speculars peak and fall | Hard ceiling at 100 IRE; pile-up at 0 IRE |
| RGB parade | Rec.709 legal | Neutrals have R, G, B aligned | Shirt or wall splits across channels |
| Vectorscope | Rec.709 | Skin on the flesh line; sat inside the box | Skin off the line; blob against the box edge |
Watch the clip play with scopes up. Generated motion will crawl: a face that sits at 55 IRE on frame one can lift as the sampler "helps" the key light. That is a generation defect, not a grade you chase.
A 20-minute grading order
Park the look. The order is the whole method. Time it once on a five-clip scene so the muscle memory exists.
Minutes 0-3: set the room. Rec.709 legal on all three scopes. Disable every LUT, every "film" preset, every automatic colour match. If the NLE applied a conversion LUT on import, take it off.
Minutes 3-8: exposure, one clip at a time, on the waveform. Lift or drop the master gain until the face sits in the same luma band as the master shot. Do not crush the floor to make it "cinematic." If the highlight region is already a flat line at 100 IRE, there is no highlight to recover. Stop and regenerate that plate with dimmer practicals and a named key that does not blow the chrome. The grade cannot invent code values the compressor already threw away.
Minutes 8-13: balance, on the parade and the vectorscope. Put a known neutral on centre. Align R, G, B on that region with offset/temperature/tint, not with a creative cube. Then check skin on the flesh line. If aligning the shirt knocks the face off the line, regenerate; do not add a second LUT.
Minutes 13-16: saturation, on the vectorscope. One global sat control. Generated clips usually need a pull, not a push. Stop when skin holds and a saturated prop (a red label, a blue wall) does not smear.
Minutes 16-20: look, last, on a layer above the matched clips. One LUT or one manual curve, one mix value, for the whole scene. If a clip breaks, the break is in exposure or balance. Do not fork the look. If you do not have a show LUT yet, a gentle S-curve and a small sat trim is enough.
Why generated clips do not match in the timeline is the assembly-side version of the same discipline: conform frame rate and resolution before you grade. Versely timelines default to 25 fps. A 24 fps insert on that timeline is a motion problem, not a colour problem.
What the scopes will not tell you
Scopes do not see extra fingers, a morphing logo, or a jacket that changes colour at the cut. Those are content defects. File them under the defect taxonomy and reroll; do not "fix" them with contrast.
Scopes also do not see compression until you stretch it. A sky that looks smooth on the monitor can already be a staircase of 8-bit steps. If the waveform shows a smooth ramp and the picture still bands after the look, the working file is too thin. Move the grade onto a 10-bit intermediate and dither; do not keep pulling the curve.
Do not grade on the editor's 480p preview. The preview is for cut, text, and timing. How that pass is charged is in editor previews and the final export. Judge Rec.709 legal range on the exported master.
When you generate the next batch, shrink this pass at the source: one model per scene, the same lighting sentence in every prompt, and no "cinematic" or "dramatic" adjectives that bake a private grade into the sample. Scopes then confirm a match instead of constructing one.
FAQ
Should I convert generated clips to log before I grade?
No. There is no scene-referred encoding in the file to recover. A "log" conversion LUT on Rec.709 display-referred footage just flattens a finished picture so you can then re-contrast it with a print LUT. That is two losses and no new information. Grade Rec.709 in Rec.709.
Do I set scopes to full range if the file looks like RGB 0-255?
Probe the file. If it is full-range RGB tagged as Rec.709, the honest fix is to interpret it as full range, then convert to legal as a delivery step, not to pretend the codes already mean 16-235. If you leave the tag wrong and set scopes to legal, blacks sit below 0 IRE and whites sit above 100 IRE, and every correction you make is compensating for a tag.
Can I skip the parade and just eye-match skin?
You can, and you will miss the shirt. Skin can look acceptable while a wall splits green-to-magenta at the cut. The parade is the cheap way to see that. Eye-matching skin is the last check, not the only one.
Why twenty minutes, not a proper session?
Twenty minutes is the budget for a short scene of already-acceptable plates. If you cannot get exposure and balance in that window, the plates are not acceptable. Extending the session usually means you started with the look, or you are trying to salvage a clipped highlight. Stop, regenerate, and spend the twenty minutes on plates that have values to move.