Your 4K Upscale Looks Worse Because You Picked the Wrong Preset, Not Because 4K Is Bad

Most people who say "AI upscaling made my video look worse" are not wrong — they are just blaming the wrong variable. They ran a 360p clip through a 4× model, got a 1440p file that looks plastic, and concluded that upscaling is a scam.

It isn't. What happened is that they applied one restoration model to content that needed four different ones. A 4K output is not a single product; it is the last step of a chain, and every preset in that chain is optimising for a different kind of damage.

This post walks through the five presets I use most, what each one actually does to pixels, and the decision rules that tell you which to reach for. If you skip to one thing, skip to the decision table at the bottom — but the reasoning matters more than the table, because presets interact.

The mistake: treating upscaling as one operation

A neural upscaler does two jobs at once:

1. Reconstruction — invent plausible detail where the source has none (edges, texture, hair strands).
2. Restoration — remove the artefacts that the previous compression already baked in (blocking, ringing, banding, mosquito noise).

These two jobs fight each other. If you push reconstruction hard on a source full of JPEG blocking, the model confidently sharpens the blocking into crisp little rectangles. That is the "worse after upscaling" phenomenon, and it is a model-selection error, not a 4K problem.

So the real question is never "which upscaler is best?" — it is "what kind of damage does this specific file have, and which model is least likely to amplify it?"

What each preset actually optimises for

Here are the five I keep in rotation, with a straight description of the trade-off each one makes. All examples below are the same subject, same source resolution, same upscale factor — only the preset differs.

Natural HD — the safe default

This is the preset that tries hardest to not look AI-processed. It reconstructs the minimum detail needed to make an image read as sharp at normal viewing distance, then stops.

• Strength: almost no halo artefacts, skin stays skin-coloured instead of going waxy, text stays readable.
• Weakness: it will refuse to invent hair strands or fine fabric weave that the source genuinely lost. If you need forensic-level detail, it under-delivers on purpose.
• Use it for: anything where realism matters more than sharpness — portraits, product photos, real estate, interview footage, family photos.

If you only ever use one preset, use this one. It fails gracefully, which is more than most models can say.

Ultra Crisp — for content that is sharp but soft

Ultra Crisp assumes the source has clean edges that were merely blurred, not destroyed. It pushes edge contrast aggressively.

• Strength: excellent on graphics, illustrations, product shots on white, UI screenshots, cosplay and costume detail, anything with hard boundaries between subject and background.
• Weakness: on noisy or heavily compressed sources it sharpens the noise too. Halos appear around high-contrast edges. Skin can go harsh.
• Use it for: content that was downscaled rather than badly encoded. If the source is clean-but-small, this is the best choice.

The distinction matters: downscaled means detail was averaged away and can be partly recovered. Badly encoded means detail was replaced by compression artefacts, and aggressive sharpening makes those artefacts louder.

True Texture — reconstruct fabric and material

True Texture is tuned for material surfaces: knitwear, fur, denim, leather, wood grain, hair. It prioritises plausible micro-texture over edge sharpness.

• Strength: clothing and surfaces stop looking like painted plastic. Fur and knit read correctly at 100% zoom.
• Weakness: it can invent texture that was never there, which is a problem if you need accuracy (medical, archival, documentary evidence). It is also slower.
• Use it for: fashion, portraits with visible clothing detail, archival photo restoration where you want the material to feel real.

The thing to understand about True Texture: it is the preset most willing to hallucinate. That is a feature for marketing imagery and a liability for anything evidentiary.

Cinema Vivid — colour and contrast, not resolution

Cinema Vivid barely touches resolution. What it changes is tone mapping — saturation, local contrast, black point.

• Strength: flat, dull, log-ish or poorly graded footage comes alive. Works well on animation and stylised content where colour separation is the point.
• Weakness: it will push an already-saturated source into clipping. Skin tones drift orange. Not a restoration preset at all.
• Use it for: social content, animation, thumbnails — anything where impact beats accuracy.

Be honest with yourself about this one. It is a colour grade wearing an upscaler's clothes. That is fine, as long as you know it.

Clear Reality — the all-rounder for real-world damage

Clear Reality tries to balance reconstruction against artefact suppression. It is the preset I use when I do not know the source's history — phone video of unknown provenance, downloaded clips, screenshots of screenshots.

• Strength: handles mixed damage (some blur, some compression, some noise) without spectacular failure in any direction.
• Weakness: it is the compromise option, so it is rarely the best at any single thing. On a clean source, Ultra Crisp beats it. On fabric, True Texture beats it.
• Use it for: heterogeneous real-world footage where you cannot inspect every frame.

Why the ordering matters more than the preset

This is the part almost every tutorial skips. Preset choice is the second decision. The first decision is what happens before the upscaler runs.

If your source has visible blocking or mosquito noise, run a denoise pass first. If you upscale first and denoise second, the upscaler has already amplified the artefacts into high-contrast detail, and the denoiser now has to fight its own model's output. You lose twice.

The practical order that works:

1. Inspect at 200–400% zoom. Identify the dominant damage: blur, blocking, noise, or banding.
2. Denoise if blocking or noise is present. Light touch — you want to remove artefacts, not detail.
3. Correct colour/exposure before upscale, not after. Upscaling an over-exposed frame bakes the clipping in.
4. Upscale with the preset matched to the content type (table below).
5. Sharpen last, and gently. A small unsharp mask after upscale recovers perceived crispness without re-introducing halos.

Decision table

The 2× rule nobody mentions

Doubling twice in sequence (360p → 720p → 1440p) almost always beats jumping straight to 4×, for two reasons: each step has less room to hallucinate, and you get a checkpoint to inspect before committing.

The cost is time. The benefit is that when step one goes wrong, you find out before you have baked the error into a 4K file you then have to redo.

What 4K actually buys you

If your source is 360p, a 4K output is not "4K quality" — it is a 360p image with 4K container dimensions. What you get is:

• Fewer visible compression artefacts at normal viewing distance (the main real benefit)
• Better perceived sharpness if the preset matched the content
• No new information. Detail that was never captured cannot be recovered, only plausibly invented.

That last point is the one that separates useful upscaling from disappointment. Set expectations at "cleaner and more presentable," not "restored to native 4K," and the results stop feeling like a scam.

If you want to try the presets on your own footage without installing anything, they run in-browser at upgolab.com — free tier included. The images above are the actual preset outputs, not marketing renders.

If this was useful, the follow-up worth reading is why temporal consistency, not resolution, is the real bottleneck in video upscaling — same channel, and it explains why a frame that looks perfect can still flicker in motion.