Guides

    Your upscale made the flicker worse

    Sharpening an upscale past a conservative setting amplifies temporal noise. Stabilize the clip first, then upscale gently so flicker dies instead of blooming.

    Versely Team8 min read

    The upscale did not fail because it was too weak. It failed because it was too eager. You sent a slightly boiling clip through a neural upscaler, cranked the sharpness, and the shimmer that was barely visible at 720p is now a sparkling mesh of pores, threads and edge halos at 1080p or 4K. Pause any frame and it looks crisp. Play it, and the surface cannot decide what it is.

    That is not a bug in the upscaler so much as the upscaler doing exactly what a sharpness control is for: boosting high-frequency residual. On a still, residual is detail. On generated video, a lot of that residual is temporal consistency error: the model re-deriving texture a little differently every frame. Sharpening past a conservative setting does not hide that error. It prints it.

    Why a sharper pass makes boiling worse

    Diffusion video re-synthesises appearance as it goes. Nothing in the sampler is a pixel tracker. Fine repeating structure (fabric weave, skin pores, foliage, hair against a bright wall) is the first place two adjacent frames disagree, because those details were never pinned down by the prompt. Practitioner write-ups on the same failure, including Picto, Hailuo and Vegavid, all land on that mechanism: the clip is individually plausible and collectively unstable.

    A generative upscaler then predicts missing high-resolution detail from that unstable source. Two consequences follow.

    It invents independently. A video upscaler is a different job from running an image upscaler on every frame, but even dedicated video models still have to invent texture. Where the source already disagrees with itself, the invention has more than one "correct" answer. Those answers crawl.

    Sharpness is a high-pass. It boosts the very band where frame-to-frame disagreement lives. On Topaz's own sharpness control, practitioners treat roughly 50 as the ceiling. Past that, the pass starts amplifying the temporal noise it was meant to hide. Versely's Topaz video upscaler does not expose that slider as a numbered field; if you are running Topaz elsewhere, keep the control in the conservative half of the range. Inside the catalog, the equivalent discipline is: pick a video upscaler, pick a target, and do not stack a second "make it crunchy" pass on top.

    The diagnostic is simple and people skip it. Play the source at full speed, then play the upscale at full speed. If the boil is already in the source, the upscaler did not cause it; it enlarged it. If the source is calm and the upscale sparkles, the pass itself is the offender, and the fix is settings and order, not another generation.

    Stabilise first, enlarge second

    The working order is not "upscale, then worry about flicker." It is the reverse.

    1. Keep the generation short. A single pass under about five seconds is the range where temporal error stays manageable. Chain a second clip rather than asking one generation to hold a long take. Error compounds with length; an upscaler will faithfully enlarge whatever the tail has already invented.
    2. Review at speed, not on the playhead. Scrubbing hides boiling. Watch the clip, then step only the frames that already looked wrong in motion.
    3. Run a temporal pass before any resolution pass. That means a deflicker or a dedicated frame interpolation stage that agrees adjacent frames with each other, not a sharpness stack. Interpolation invents in-between motion; it does not add detail, which is exactly what you want here. You are buying agreement, not crunch.
    4. Then upscale, conservatively. One video upscaler, one target. ByteDance's video upscaler exposes 1080p, 2K and 4K as the target list; start at 1080p even if the deliverable is 4K, inspect, then step up. Two modest steps beat one aggressive jump, because each jump is another chance to invent texture that does not exist in the source.
    5. Do not sharpen to taste on top. If the result is still soft, the source never had the information. Soft and stable ships. Sharp and boiling does not.

    Cleanup lives in front of finishing for the same reason. The editor's cleanup tools belong before upscaling to 4K, not after. Going straight to 4K enlarges every shimmering strand faithfully. The upscaler has no way to know the detail it is amplifying was inconsistent to begin with.

    Stage Job What "done" looks like
    Generate Motion, framing, identity Shot works at native size; boil is minor
    Temporal Agreement across frames Stationary texture holds still when played
    Upscale Resolution only Same clip, bigger, no new crawl
    Grade / export Look Colour, not structure

    If you reverse rows two and three, row three becomes a flicker amplifier. That is the whole article.

    What to inspect after the upscale

    Check the surfaces that fail first, at 100 percent, in motion:

    • Skin. Pores that pop in and out are sharpened temporal noise, not "detail the model finally found."
    • Fabric. Weave and stitching that crawl, even when the garment is not moving.
    • Hard edges. A halo that breathes around hair, type, or a product silhouette.
    • Repeating structure. Brick, mesh, foliage, crowd texture. Reconstruction likes to regularise these, then disagree about the regularisation every frame.

    If any of those crawl and the source did not, drop the target resolution one rung and run the upscale again. If they crawl in the source too, do not upscale at all until a temporal pass or a shorter regeneration has calmed it. Whether to upscale the render or generate at full resolution is a different decision; it is about whether the upscaler is allowed to invent information (labels, distant faces). This post is about the case where you already decided to upscale, and the pass made the clip less watchable.

    A grade will not save you. Colour correction moves values; it does not make frame twenty-one agree with frame twenty. Sharpening in the grade is the same high-pass with a different label. Leave the finishing LUT for after the texture holds still.

    A conservative recipe that actually ships

    Use this when a keeper take is slightly soft and slightly alive in the wrong way.

    Generation. One subject, one motion idea, three to five seconds. Busy backgrounds and loose hair make boiling cheaper to create and more expensive to hide; simplify the frame if you can still make the shot.

    Temporal. Deflicker or interpolate only as much as you need for adjacent frames to agree. Do not interpolate in order to manufacture slow motion on a clip that already shimmers; you will smear the boil into a longer smear.

    Upscale. One dedicated video model, not an image model run per frame. ByteDance and Topaz are video-native. Pick the lowest target that satisfies the delivery spec. Inspect. Step up only if the first pass held.

    Confirm on a preview, then export once. The editor's 480p preview pass runs at no credit cost and carries a short per-user cooldown; the final export is charged once regardless of clip count. Confirm the boil is gone at preview size before you pay for the finishing render. Preview will not show 4K texture, but it will show whether the surface is still crawling, which is the question you are answering.

    If a localised patch is the only offender (a sleeve, a wall, a patch of hair), do not re-upscale the whole shot at a higher sharpness. Mask that region and repair it, then upscale the repaired clip. Re-rolling the generation to chase a texture problem you introduced in finishing is how a one-hour job becomes an afternoon.

    FAQ

    Is a higher-resolution generation immune to this?

    No. Native 4K still re-synthesises texture per frame. It can be more stable than a soft source plus an aggressive upscale, because you are not asking a second model to invent pores on top of unstable pores. It is not a flicker-free format. Length, motion and repeating detail still decide whether the surface holds.

    Can I fix boiling by generating at a higher frame rate?

    Usually the opposite. More frames over the same duration is more opportunities for adjacent frames to disagree. Generated footage sits most comfortably in the 24-to-30 fps band, and video generations here default to 25 fps, inside it. If a delivery needs 50 or 60, interpolate after the clip is temporally calm, not before.

    Why does the paused frame look fine?

    Because boiling is not a still-image defect. Each frame is a plausible high-resolution picture. The error is that the plausible pictures do not share a texture identity. Review at full speed. If you only scrub, you will keep shipping the problem.

    When should I skip the upscale entirely?

    When the detail that matters is information rather than texture: small text, a logo, a face that is small in frame. An upscaler will invent letter-shaped marks and reconstructed eyes, and sharpening will make the invention louder. Generate those shots at the delivery resolution, or composite the type. For everything else, stabilise, then upscale gently, and stop at the first pass that holds still.