The order of finishing passes: clean, then scale
Going straight to 4K amplifies artifacts you could have removed cheaply. The pass order that works, and what each pass can and cannot fix.
An upscaler does not know which parts of your frame are the subject and which are mistakes. It enlarges both. Run a shimmering, slightly mismatched sequence through a 4K pass and you get a shimmering, slightly mismatched sequence at four times the pixel count, at four times the render time, and with the artifacts now large enough that nobody can miss them.
The standard finishing order exists to stop that: clean first, then scale. It is one line of advice and it saves more time than any single tool choice, because every pass after the first is operating on whatever the first pass left behind.
The order
Five passes, in this sequence. Each one assumes the ones above it have already run.
| # | Pass | What it is for | What it cannot do |
|---|---|---|---|
| 1 | Deflicker / clean | Temporal instability: shimmer, crawling texture, frame-to-frame noise | Fix geometry, hands, or anything structurally wrong |
| 2 | Colour match | Exposure and white-balance drift between clips | Rescue a blown highlight or a crushed black |
| 3 | Scale to 1080 | Get every clip to one working resolution | Add detail that was never rendered |
| 4 | Scale to 4K | Only if the destination genuinely uses it | Recover the pass you skipped at step 1 |
| 5 | Sharpen | Perceived crispness on a finished, stable image | Undo the halos it creates if overdone |
The ordering logic is the same at every step: each pass amplifies whatever is already in the frame, so anything you can remove cheaply should be removed before something expensive multiplies it. Deflicker is cheap and its input is the raw generation. Sharpening is the most amplifying operation of the five, which is why it goes last and why "sharpen last" is worth treating as a rule rather than a preference.
Pass 1: clean before anything
Temporal artifacts are the ones to attack first, because they are the ones an upscaler handles worst. A spatial upscaler works largely frame by frame; a texture that is unstable between frames gives it a different problem to solve on every frame, and it solves each one slightly differently, which makes the instability worse rather than better. Shimmer that was mildly annoying at 720 becomes the dominant feature of the shot at 4K.
Deflicker, then. This is also the moment to make the harder call: some things are not artifacts, they are the shot being wrong, and no finishing pass fixes them. Hands with the wrong number of fingers, text that is not real letterforms, a limb that detaches during an occlusion. Those are generation problems with generation-side fixes such as inpainting the region or regenerating the shot, and the catalogue of which is which sits in why AI video artifacts happen.
Deciding this early is the point. A reroll at step 1 costs one generation. A reroll discovered at step 4 costs the generation plus everything you did in between.
Pass 2: colour match every clip
Colour matching across clips is not a look decision, it is a continuity repair, and it belongs before resolution work rather than after. Generated clips drift in exposure and white balance across a chained sequence, and matching them at assembly is the direct counter. On any multi-clip piece it is non-optional.
Do it before scaling for two reasons. First, grading a 1080 timeline is faster than grading a 4K one and the decisions are identical. Second, an upscaler trained on natural imagery behaves differently on a clip that is a quarter-stop hot than on one that is correctly exposed, so matching first means the scale pass sees consistent input across the sequence instead of a moving target.
Know the limits before you start. A grade moves the relationship between tones; it does not invent information that was never rendered. Detail lost in a blown highlight is gone, and pushing a crushed shadow open surfaces noise rather than shadow detail. What a grade can and cannot rescue is the honest version of that boundary, and it is the difference between an hour well spent and an hour spent avoiding a reroll you were always going to need.
Passes 3 to 5: scale, then sharpen last
Get everything to 1080 first. Standardise resolution across the sequence before going any higher. Mixed-resolution timelines force a scale on some clips and not others at export time, and a scale you did not choose is a scale you did not control.
1080 is also, for a lot of work, the end of the road rather than a waypoint. Social feeds compress hard, and the difference between a well-cleaned 1080 master and a 4K one is frequently invisible after a platform's transcode. Whether to go further is a real decision with a real cost attached: when a 4K export is worth the credits and what 4K costs against 1080 are the places to check before assuming upward.
Then 4K, if the destination pays for it. Step up rather than jumping. A common pattern is a fast 2× pass to get most of the way, followed by a refinement pass that fixes what the fast one approximated. Two moderate steps give each pass closer input to work from than one large step does.
Model choice matters here, and it is not the obvious ranking. Purpose-built video super-resolution models aimed at generative footage, of which SeedVR2 is the most cited and FlashVSR the faster alternative, hold temporal consistency well on AI-native material. Topaz Video AI remains the reference for live-action restoration, resolutions above 4K and HDR finishing, but it can be too inventive on generated footage, hallucinating detail that was never there. The right tool depends on what the source is, not on which is generally better.
Settle one question before this pass: whether you should be upscaling at all rather than regenerating at the higher resolution. Some models render natively at high resolution, and upscaling the render versus generating at full resolution takes the tradeoff apart. If you do upscale, upscaling video to 4K is the one-step version, and the upscaling glossary entry covers what the operation does and does not add.
Sharpen last, and less than you want. Sharpening increases local contrast at edges. It does not add information. Applied to a clean, stable, correctly scaled image it makes the picture read as crisper. Applied earlier in the chain it hands every subsequent pass an exaggerated version of whatever was already there, including the noise.
Two failure modes to watch for. Halos: a bright fringe on the dark side of a high-contrast edge, most visible against sky or a plain wall. And re-introduced shimmer: sharpening amplifies frame-to-frame differences, so a sequence successfully deflickered at step 1 can start crawling again if you sharpen hard at step 5. If that happens, the answer is less sharpening, not another deflicker pass.
Running the whole thing
The order collapses to six operations:
- Reroll anything structurally broken. Do not finish a shot you are going to replace.
- Deflicker and clean at source resolution.
- Colour match every clip to a chosen hero clip.
- Assemble at 1080. Lock the cut.
- Preview.
preview: truegives a 480p pass at no credit cost with a short per-user cooldown, and the final export is charged once regardless of how many clips are on the timeline. Timing, drift and structural problems all survive the downscale; fine detail obviously does not, which is exactly why this check belongs before the resolution work rather than after. - Export at 1080, or step to 4K and sharpen only if the destination uses it.
Everything from step 3 onward lives on one EDL-based timeline in the AI video editor, which means the sequence is a re-renderable object rather than a stack of intermediate files. That matters more than it sounds for a pass order like this one, because the usual reason people skip step 1 is that going back to it would mean redoing everything after it.
FAQ
Why does going straight to 4K make things worse?
Because an upscaler amplifies whatever it is given, without distinguishing subject from defect. Temporal artifacts suffer most: a spatial upscaler resolves each frame slightly differently, so an unstable texture becomes more unstable at higher resolution rather than less. Removing the instability first is both cheaper and more effective than trying to out-resolve it.
Can I skip the 1080 step and go from source straight to 4K?
You can, and for a single clean clip it is often fine. The reason the intermediate step exists is sequences: getting every clip to one resolution before scaling means the scale pass treats them identically, and it lets you lock the cut and grade at a resolution that is fast to work in. Two moderate steps also generally beat one large step, because each pass has closer input to work from.
Does sharpening ever belong earlier in the chain?
Not in this order. Sharpening is the most amplifying operation of the five, so anything it runs before will magnify what it did. The one adjacent case is a capture-style sharpen baked into an upscaling model's own output, which you cannot separate out — in that situation, treat the model's result as already sharpened and add little or nothing on top.
How do I know whether a defect is a finishing problem or a generation problem?
Ask whether it is stable. Temporal problems — shimmer, crawling texture, flicker — are finishing problems and a deflicker pass is the right tool. Structural problems that persist identically across frames — wrong hand geometry, garbled text, a face that is not the right face — are generation problems, and no amount of cleaning, grading or scaling touches them. The second category needs an inpaint, a reference change, or a reroll, and finding out at step 1 is much cheaper than finding out at step 4.