What debanding settings should I use before AI upscaling video to 4K?

The Short Answer: Apply Debanding Before Upscaling, Not After

If you are preparing a video for AI upscaling to 4K, the correct order of operations is to run a debanding pass first, then feed the cleaned file into your upscaler. Banding — the visible stair-step rings in gradients like skies, shadows, and dark scenes — is compression damage baked into the source pixels. When an AI upscaler processes banded footage, it treats those hard edges as real image structure and sharpens, enhances, or even amplifies them at the higher resolution. The result is a 4K file where the banding is not only preserved but made more obvious, because every gradient step now spans more pixels on screen.

Also worth reading: What are the best AI upscaling settings for gameplay in 2026? · SeedVR3 ComfyUI 4K settings explained: how do you configure SeedVR3 for 4K upscaling in ComfyUI? · What are the best HDMI 2.1 settings for a 4K TV to get the highest quality picture and AI upscaling performance?

The practical settings that work for most content are conservative ones. In tools like AviSynth with f3kdb (Flash3kyuu Deband), a typical starting point is a range of 16–24 pixels, a threshold of around 48–64 (on the 0–255 scale) for luma, and a blur strength of roughly 20–30, applied in 16-bit processing depth. In ffmpeg's gradfun or deband filters, start with default thresholds and raise them only if banding remains visible after a test encode. The goal is to smooth the gradient steps just enough that the upscaler sees continuous tonal transitions — not to blur the entire frame. Over-debanding destroys fine film grain and texture, which modern upscalers actually rely on as detail cues, so restraint matters more than aggression.

Why Banding Breaks AI Upscalers Specifically

AI upscaling models such as NVIDIA's RTX Video Super Resolution (released for RTX 30 and 40 series GPUs in early 2023) and commercial tools like Topaz Video AI are trained on paired low- and high-resolution imagery. They learn statistical patterns about what edges, textures, and gradients should look like when enlarged. A banded gradient violates those learned assumptions: the model interprets each contour line between two flat color steps as a genuine edge worth enhancing. On a 4K display — especially a large OLED or a high-brightness monitor like the LG UltraGear Evo 39GX950B reviewed by TFTCentral, which can push highlights hard enough to expose gradient flaws — those enhanced contour lines become clearly visible artifacts rather than subtle ones.

The problem compounds because most sources destined for upscaling are heavily compressed. Streaming rips, old Blu-ray encodes, and DVD captures were often encoded at 8-bit color depth with aggressive quantization, which is precisely the combination that produces banding in dark scenes and skies. A 10-bit HEVC re-encode of a banded source does not fix anything; the extra bit depth simply preserves the existing steps with more precision. Debanding before upscaling addresses the root cause by reconstructing plausible intermediate tones, giving the neural network clean gradient data to work from.

There is also a temporal dimension. Single-frame debanding can produce flickering if the smoothing pattern shifts independently between frames, because the human eye is extremely sensitive to noise movement in flat areas. Tools like f3kdb support temporal radius settings (typically 1–2 frames) that average the debanding across neighboring frames, stabilizing the result. For AI upscaling workflows, enabling temporal processing is strongly recommended whenever the source has any motion in gradient regions, which is nearly all real-world footage.

Recommended Starting Settings by Tool

The exact parameters depend on your software chain, but the underlying logic is identical everywhere: detect gradient regions, add dithering or controlled blur to break up the steps, and protect textured areas from being smoothed. Below is a comparison of the most common approaches used in pre-upscaling pipelines.

For f3kdb, a widely used configuration looks like this conceptually: input converted to 16-bit, range set to 16, y/cb/cr thresholds at 64/64/64, grain extraction enabled with a grain value of 32–48 so original film texture is sampled and reapplied after smoothing, and output dithered back down. This last point deserves emphasis — always dither when converting back to 8-bit for delivery, because naive bit-depth reduction reintroduces exactly the banding you just removed. If your final delivery is 10-bit HEVC, you can skip the downward conversion entirely and retain the full benefit.

For ffmpeg users, the deband filter offers more control than gradfun: parameters include 1thr through 4thr (thresholds per plane, defaults around 0.012), range (default 16), blur direction, and coupling. A reasonable starting command uses thr values of 0.02–0.03 for visibly banded dark content and 0.01 for mild cases, with range left at 16. Test on the worst 30 seconds of your footage — usually a fade-to-black or a night scene — before committing to a full batch render.

How to Diagnose Whether Your Source Needs Debanding

Not every source needs treatment, and applying debanding indiscriminately wastes render time and softens detail. The reliable diagnostic method is visual inspection under conditions that reveal banding: view the footage fullscreen on a quality display in a dim room, and scrub specifically through gradient-heavy content — cloudless skies, fog, smoke, dim interiors, lens flares, and fade transitions. Dark scenes are the classic failure zone because 8-bit video allocates fewer code values to shadow ranges; a night sky might have only 20–30 distinct levels across what should be a smooth falloff.

A more objective test is to examine a histogram or waveform scope of a suspect region. Flat plateaus separated by sharp jumps in the histogram indicate quantization steps. You can also pause on a gradient and zoom to 200–400%: true film grain appears as random fine noise, while banding appears as concentric or parallel contour lines with flat areas between them. If you see contour lines even at normal viewing distance on a large screen, the source needs debanding. If banding is only visible when you press your nose against the monitor, leave it alone — the upscaler will likely handle it acceptably, and aggressive smoothing will cost you more than it gains.

Bit depth of the source is a useful prior but not a verdict. A well-encoded 10-bit source rarely bands; an 8-bit source from a high-bitrate master often does not either. Conversely, low-bitrate 8-bit streams (typical of older web downloads below roughly 3–4 Mbps for 1080p) almost always show banding in gradients. Judge by inspection, not by spec sheet.

Common Mistakes That Ruin Pre-Upscaling Debanding

The single most frequent error is over-thresholding. Cranking thresholds to 100+ or using maximum blur strengths flattens everything that resembles a gradient, including skin textures, fabric weave, foliage, and water surfaces. The AI upscaler then invents detail to fill the void, producing the plasticky, waxy look that gives AI enhancement a bad reputation. Keep thresholds at the minimum value that eliminates visible contours on your test clips, and always enable grain preservation or re-graining when working with film-sourced material.

The second common mistake is wrong ordering in the pipeline. Some users upscale first and deband afterward, reasoning that the larger canvas makes smoothing easier. This is backwards: once the upscaler has sharpened the contour lines into defined edges, no amount of post-smoothing removes them without also destroying legitimate detail. Others apply denoising before debanding with overly strong settings, stripping the very grain that helps both the debander and the upscaler distinguish texture from artifact. If your source is noisy and banded, use a light denoise (or a grain-aware denoiser) followed by debanding with grain retention, then upscale.

A third mistake involves color space handling. Running debanding filters on the wrong plane or in RGB when the tool expects YUV (or vice versa) can shift colors or miss chroma banding entirely. Chroma banding — visible as blotchy color transitions in red sunsets or blue skies — requires the chroma planes to be processed too, which f3kdb handles via its cb/cr thresholds but some simpler filters ignore. Finally, avoid double compression: encode your debanded intermediate at a high bitrate or losslessly (or at minimum 10-bit CRF 14–18 x265), because a sloppy intermediate encode will re-band the footage before the upscaler ever sees it.

Debanding Versus Alternatives: Dithering, Denoising, and Doing Nothing

Debanding is one of several tools that address gradient quality, and knowing when each applies saves considerable time. Dithering adds structured random noise during bit-depth conversion to perceptually smooth gradients; it prevents new banding from forming but cannot remove banding already present in the source. If your pipeline converts 16-bit processed intermediates down to 8-bit delivery files, dithering is mandatory regardless of whether you deband. If your source is already banded, dithering alone is insufficient.

Denoising overlaps with debanding but targets different damage. Noise is random variation per pixel; banding is systematic quantization structure. Strong denoisers incidentally reduce banding by flattening the flat areas between contours, but they do nothing to bridge the steps themselves, and they sacrifice texture. Dedicated debanders explicitly model the gradient and reconstruct intermediate values, which is why they outperform denoisers on this specific defect. The two operations are compatible and often sequenced together for damaged archival material.

Doing nothing is sometimes the right call. Modern hardware solutions like NVIDIA RTX Video Super Resolution operate in real time on playback and include their own artifact handling; for casual viewing of mildly banded content on a mid-size screen, the improvement from a full offline deband-and-upscale pipeline may not justify hours of rendering. The calculus changes for archival projects, large-screen viewing (65 inches and up), OLED displays with perfect blacks that expose shadow banding, or content you will watch repeatedly — there, the upfront effort pays off permanently in the master file.

## When to Act: Pipeline Position and Timing

Debanding belongs immediately after any decoding and cropping steps and before denoising decisions are finalized, scaling, and encoding. Concretely, a robust workflow reads: decode source → trim/crop → light denoise if needed → deband with grain preservation → optional detail restoration → AI upscale to 4K → encode final deliverable in 10-bit HEVC or AV1. Placing debanding before the upscale is non-negotiable for the reasons covered above; placing it before or after light denoising is a judgment call depending on how noisy the source is, since heavy noise masks banding detection and should be reduced first.

Timing-wise, budget for iteration. A careful operator tests three to five parameter combinations on representative 15–30 second clips before launching a full render. With f3kdb running at 16-bit depth, expect processing speeds of roughly 5–20 fps on a modern desktop CPU depending on resolution and temporal radius — a 90-minute film can take several hours for the deband pass alone, before the AI upscale, which on an RTX 40-series GPU typically runs faster than real time for 1080p-to-4K with models like those in Topaz Video AI or comparable local tools. Plan accordingly: the deband pass is cheap relative to the upscale, so getting its settings right through short tests costs minutes, while discovering banding survived after a 12-hour full render costs a day.

Revisit settings per source, not per project. A batch containing mixed sources — a clean Blu-ray rip alongside a rough web download — should be split into groups with different deband intensities, because uniform settings will over-treat the clean source and under-treat the damaged one. Automation-minded users can script threshold selection based on measured histogram plateau counts, but manual per-source tuning remains the standard practice among restoration communities as of 2026.

Cost Considerations and Practical Value

The software side of debanding is essentially free. f3kdb, VapourSynth, AviSynth+, and ffmpeg are all open-source, and every major operating system supports them. The real costs are time and, if you lack suitable hardware, compute. AI upscaling itself ranges from free options (RTX Video Super Resolution for playback on supported GeForce cards, released February 2023; open-source inference tools) to paid suites like Topaz Video AI at $299 for a perpetual license with a year of updates. None of these prices change based on whether you deband, so the marginal cost of proper debanding is purely the extra rendering hours and learning curve.

Whether that investment is worthwhile depends on your display and content. On a 27-inch desktop monitor viewed at arm's length, residual banding after a decent AI upscale is often barely noticeable. On a 77-inch OLED in a home theater, or on high-end gaming displays like the 39-inch LG UltraGear Evo 39GX950B that TFTCentral tested — panels capable of extreme brightness and deep blacks that mercilessly expose gradient defects — untreated banding is readily visible and undermines an otherwise excellent 4K presentation. For anyone building a permanent library of upscaled classics, the few hours spent on debanding per title is among the highest-return steps in the entire pipeline, second only to choosing a good upscaling model in the first place.