Topaz Upscale: Pick the Right One of 7 Topaz Models
There isn't one Topaz upscale. Here are the seven Topaz models on each::labs sorted by job, from photos and tiny thumbnails to old footage and slow motion, with the settings that matter for each.
By Eftal Yurtseven · Nov 18, 2026 · 6 min read

Everyone who searches "topaz upscale" wants the same thing: a bigger, sharper version of something they already have. The trap is thinking there's one Topaz upscale. There isn't. On each::labs there are seven Topaz models, and the wrong pick either smears a face into plastic or spends ten minutes sharpening noise that should have been removed.
This guide sorts them by job. Photos and graphics go one way, old or compressed footage another, slow motion a third. For each one you get the settings that actually change the result, taken from the each::labs request schemas, so you can run a Topaz upscale on the first try instead of the fifth.
The Seven Topaz Upscale Models at a Glance
Split them by input first, because nothing else matters until you do.
For still images you have three. Topaz Image Upscale is the all-rounder with eight model presets inside it. Gigapixel Standard 2 is the balanced enlarger with fine controls for noise, compression and faces. Wonder 3.5 is the one for genuinely low-resolution images that need detail rebuilt, not just stretched.
For video you have four. Topaz Upscale Video is the quick multiplier. Proteus 1 is the hands-on cleaner for low to medium quality footage. Starlight Precise 2.6 is the restoration model for faces, textures and text. And Apollo doesn't add pixels at all: it adds frames, for slow motion and frame rate conversion.

Topaz Upscale for Photos: Start With Image Upscale
If you only learn one image endpoint, make it Topaz Image Upscale. It takes an upscale_factor from 1 to 4 (default 2) and a model field with eight presets: Standard V2, Standard V1, High Fidelity V2, High Fidelity V1, Graphics, Low Resolution, CG Art & Game Assets and Text Refinement.
The preset is the decision. Standard V2 is the default and a safe start for ordinary photos. High Fidelity is the one to try when the source is already decent and you want it kept faithful rather than reinterpreted. Graphics and CG Art & Game Assets exist because flat colors and hard edges break photo models. Low Resolution is for tiny sources. Text Refinement is for screenshots, signage and product labels, where a photo model will happily turn letters into soup.
Faces get their own switch. face_enhancement is on by default here, with strength at 0.8 and creativity at 0. Leave creativity at zero for real people: higher values give the model more room to invent features, which is fine for a game character and wrong for a client's headshot. subject_detection (All, Foreground or Background) decides where the enhancement lands.

Gigapixel Standard 2: When You Need Fine Control
Gigapixel Standard 2 is the enlarger for people who want to set every dial. Instead of a factor, you give it a target output_width or output_height up to 32,000 pixels; set one and it scales the other in proportion. The input can be JPEG or PNG up to 512 megapixels, and output can be JPEG, PNG or TIFF, which matters if the file is going to print.
Four sliders from 0 to 1 do the cleaning: denoise, fix_compression, sharpen and strength. The schema warns that higher strength can look less realistic, and that matches what you'd expect from any upscaler. Fix the source problem first (compression blocks on a web JPEG, noise on a night shot), then add only as much sharpening as the result needs. Face recovery is opt-in here, and when you turn it on you have to set its strength and creativity yourself.
Wonder 3.5 for Images That Were Never Big
Some images aren't soft, they're small. A thumbnail, an old avatar, a product shot pulled from a marketplace listing. Wonder 3.5 is described on each::labs as the model that turns low-resolution images into detailed, realistic upscaled outputs, and its controls are deliberately few.
enhancement_strength is low, medium or high, and the default is high. That default is aggressive on purpose. If the result looks too clean or too invented, drop to medium before you touch anything else. Wonder also has an optional grain stage (grain, with density, strength, size and a silver, gaussian or grey model). It sounds like a gimmick until you upscale a film photo and realize the grain is what made it look real. Add a little back and the "AI smooth" look disappears.
Searchers know this model by name: "topaz wonder" and "topaz wonder 3.5" both show up in our Search Console. If your source is tiny, it's usually the right first try.

Topaz Upscale for Video: Pick by What Is Wrong
Video models are where people waste the most time, because each one fixes a different problem. Ask what's wrong with the footage before you ask how big it should be.
It's fine, just small. Use Topaz Upscale Video. It has three settings that matter: upscale_factor (default 2), an optional target_fps, and h264_output if you need the most compatible file. Nothing to tune, nothing to break.
It's noisy, compressed or soft. Use Proteus 1. It defaults to Auto, where Topaz estimates the enhancement parameters itself, and that's the right first pass. Switch auto to Manual when you want the sliders: noise, compression, details, halo, blur and preblur, each from -1 to 1. There's a focus_fix_level (None, Normal, Strong) for blurred subjects and a recover_original_detail_value that mixes source detail back in when the cleaned result looks too polished. Proteus also asks for the source facts up front, including input_width and input_height, so have them ready.
It's old, and faces or text matter. Use Starlight Precise 2.6, the restoration model for faces, textures and text clarity with temporal consistency. Its one creative dial is sharpness from 1 to 5, and the default is the maximum. Lower it for skin. We wrote a full Starlight Precise 2.6 guide if restoration is your main job.
Apollo: The Topaz Model That Adds Time
Apollo is the odd one out, and the most fun. It doesn't sharpen anything. It interpolates new frames, so you can turn ordinary footage into smooth slow motion or convert between frame rates.
Two settings carry it. slowmo is the slow motion factor from 1 to 16 (default 2), and fps is the target output frame rate from 15 to 240 (default 60). There's also a duplicate switch that finds and removes repeated frames before interpolating, which is exactly what you want on screen recordings or footage that was converted badly once already. For AI-generated clips, Apollo is a neat way to stretch a 5-second hero moment into a slow reveal without regenerating it.

Build a Topaz Upscale Workflow That Holds Up
Order matters more than any single setting. For video, the reliable chain is clean first, enlarge second, retime last: Proteus 1 or Starlight Precise 2.6 to fix the picture, then Apollo if you need slow motion. Retiming a noisy clip interpolates the noise too.
For images, upscale before you edit, not after, so every later step works on clean pixels. And for anything you generate with an image model or video model, put the upscaler at the end of the same each::labs flow, so every output arrives at delivery size without a manual round trip.
Faces in old family photos are a special case. If the face is the only thing that matters, our GFPGAN face restoration guide covers a model built only for that job, and you can upscale the rest with Topaz afterwards.
Frequently Asked Questions
Which Topaz upscale model should I use for photos?
Start with Topaz Image Upscale on the Standard V2 preset at 2x. Move to Wonder 3.5 when the source is genuinely tiny, to Gigapixel Standard 2 when you need an exact output size or print formats, and to the Text Refinement or Graphics presets when the image is mostly type or flat art.
How big can a Topaz upscale go?
For images, Gigapixel Standard 2 and Wonder 3.5 accept target dimensions up to 32,000 pixels on a side, while Image Upscale works by factor, from 1 to 4. For video, Proteus 1, Starlight Precise 2.6 and Apollo take output dimensions up to 8,192 pixels, with 1920x1080 as the default.
Can Topaz make AI-generated video look sharper?
Yes, and it's one of the most common uses. Upscale Video is the quick route for clips that are clean but small. Starlight Precise 2.6 is the better pick when faces or on-screen text look mushy, and Apollo can add smooth slow motion to a finished clip.
Why does my upscaled face look fake?
Usually too much creativity or sharpening. Set face enhancement creativity to 0, lower the strength, and on Starlight Precise 2.6 bring sharpness down from 5. A light touch of grain on Wonder 3.5 also hides the over-smooth look.



