ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›ARTICLES›HOW TO UPSCALE AN AI IMAGE: WHICH METHOD TO USE
ARTICLES · HOW-TOBEGINNER

How to Upscale an AI Image: Which Method to Use and What the Settings Mean

Six ways to upscale an AI image, from generating bigger to open-source tools, and what faithful and creative settings do. Check faces and text afterwards.

READ12 min
WORDS2,322
SECTIONS9
TYPEHOW-TO
CHECKED27 OCT 26

To upscale an AI image, run it through an upscaler that makes it bigger and sharper, and choose the kind by what you want back: a faithful upscaler if the picture must stay the same, a creative one if you are happy for it to add new detail. First check whether you need to upscale at all, because some image tools can now make a 4K picture directly. And remember that the extra detail is predicted, so look closely at faces, hands and any text afterwards.

Want words to paste instead of steps? Use the ready-made upscale prompts: a universal prompt, a “change nothing, just sharper” prompt, and ones for old photos and drawings.

TL;DR — THE SHORT VERSION
  • Try a bigger size first. Some tools can generate at 2K or 4K, so no upscale is needed.
  • Pick faithful or creative. Every upscaler has this dial, under different names: Subtle or Creative, Resemblance, Creativity, Strength.
  • Start low and upscale once. Each extra pass adds more of the AI’s guesses.
  • Zoom to full size on faces, hands and lettering before you use the result, and keep the original.
◈ IN PLAIN TERMS

Enlarging a photo the ordinary way is like blowing up a balloon with a drawing on it: the drawing gets bigger and blurrier. AI upscaling is like handing the small drawing to an artist who redraws it large. A careful artist copies what is there. A creative artist adds what they think should be there. The settings on an upscaler decide which artist you get.

How do you upscale an AI image? The six ways

The plain point: there are six common routes, from “generate it bigger” to “run your own software”. The table lists what each maker says its own tool does. It does not rank them.

Six ways to make an AI image bigger, from each maker’s own pages (read 27 Oct 2026)
WayWhat it isControl you getWatch for
1. Generate biggerLike: asking the image tool for a larger picture in the first place. Google’s Gemini models and OpenAI’s image API can do this.Set: a size, such as 2K or 4K.Limit: some models stop at 1K; OpenAI marks sizes above 2560x1440 as experimental.
2. The image tool’s own upscalerLike: Midjourney’s Subtle and Creative upscalers.Choose: Subtle or Creative. Both double the size.Limit: HD images cannot be upscaled further.
3. A dedicated upscalerLike: Topaz Gigapixel, or Krea’s Enhancer.Choose: a model, a scale factor and how much to invent.Limit: the creative settings can change faces and textures.
4. An editor with generative upscaleLike: Adobe Photoshop’s Generative Upscale.Choose: a scale and one of three models (Adobe’s options are quoted below).Limit: each model has its own size cap.
5. A chat tool and a promptLike: ChatGPT or Gemini with the picture attached and an instruction.Control: only your words.Limit: no strength slider, so it may redraw more than you wanted.
6. Open-source softwareLike: Real-ESRGAN, or the Stable Diffusion upscaler in Hugging Face’s Diffusers.Set: options in code, such as a noise level.Limit: you run it yourself.

Do you need to upscale at all?

The plain point: not always. Some tools can make a large image directly, which is often cleaner than enlarging a small one afterwards.

Google says its Gemini 3 image models “generate 1K images by default but can also output 2K, 4K”, and for Nano Banana 2.1 a square 4K picture is 4096x4096 pixels.1 The smallest, Nano Banana 2 Lite, stops at 1K, so it needs an upscaler.1

OpenAI’s image API takes custom sizes too: “neither edge may exceed 3840 pixels”, and “Resolutions above 2560x1440 are experimental.”2 That is the developer API, not the ChatGPT app.

Midjourney works differently. In version 8.2, standard images start at 1024 x 1024 pixels and the Creative and Subtle upscalers take them to 2048 x 2048. HD images arrive at 2048 x 2048 and “currently cannot be upscaled further.”3

How big a picture needs to be for a screen or print is on aspect ratio, resolution and format. If the picture already looks fake, size will not fix it.

Faithful or creative: which kind of upscale do you want?

The plain point: every upscaler offers a choice between keeping your picture and improving it with new detail. The names differ; the choice is the same.

  • Midjourney: “the Subtle upscaler keeps changes to a minimum, preserving the look of your initial image as closely as possible,” while Creative “will change your image a bit as it adds new details.” It adds that Creative can sometimes help with “awkward hand positions or odd facial expressions.”3
  • Topaz Gigapixel: its Redefine model is “for use with low quality or AI generated images”, and Topaz says “You can prioritize upscaling for either realistic fidelity or creative distinction.” Redefine realistic has a None or Subtle setting; Redefine creative has Low, Medium, High and Max creativity.4 Its Recover model “brings the best upscaling fidelity for old and low quality photos.”4
  • Photoshop: Adobe lists three models, “Each model delivers a distinct type of enhancement”: Firefly Upscaler restores low-resolution images, Topaz Gigapixel “Preserves existing detail”, and Topaz Bloom “Adds new, creative detail.”5 You choose 2x or 4x.5
  • Krea: its table has an “Upscale V1” model described as a “Fast, neutral upscaler that preserves existing detail well”, and a “Krea Enhance” model described as a “Generative upscaler suited for adding detail and improving lower-resolution AI images.”6
  • Stable Diffusion upscaler (open source): Hugging Face describes it as a “Pipeline for text-guided image super-resolution” that raises resolution “by a factor of 4”, with a noise_level where “A higher noise_level adds more noise to the input, increasing variation but reducing fidelity to the source image.”7
TAKEAWAY

Whatever the tool calls it, the dial runs from “copy my picture” to “improve my picture”. Start at the copy end. Move up only if the result looks flat.

What do the upscaler settings mean?

The plain point: most settings are one dial in disguise: how much the AI may make up. Krea’s guide is the clearest about it.

Krea says Strength “Controls how much additional detail the AI invents,” Resemblance “Higher values keep the output closer to the original image,” and for Topaz Generative, Texture sets how much new detail appears: “A lower value upscales existing detail; a higher value generates additional texture”.6

Krea also warns that higher is not always better: “Upscaling a painting with high texture will fill in additional brushstrokes that were not in the original,” and on photographs, high sharpness and clarity can add “visible pores and fine lines to skin that was originally clean.” Its advice is to “start conservatively and increase settings only until you reach the result you want.”6

Two more settings are about size, not style. The scale factor (2x, 4x, 8x) multiplies width and height, so Krea notes a 2x upscale gives an image 4 times the pixels and a 4x gives 16 times. It also says that for large prints, “upscaling in steps -- 2x first, then 4x -- tends to produce more controlled results than jumping directly to 8x.”6 And a prompt or image description, where offered, tells the tool what the picture shows. Krea’s tip is to “Describe the final image you want to see, not the change you want made.”6

How do you upscale an AI image, step by step?

The plain point: six steps, in order. This is this site’s own routine, built on the advice quoted above.

  1. Start from your best version. Upscale the picture you would be happy to use at the smaller size. An upscaler enlarges mistakes too.
  2. Check the size first. If the tool can generate bigger, skip the upscale.
  3. Pick the faithful setting first. Subtle in Midjourney, Recover or Redefine realistic in Gigapixel, a high Resemblance in Krea, a low noise level in the open-source pipeline.
  4. Upscale once, by 2x. Go further only if you need the size.
  5. Look at it full size. Check faces, hands, small objects and any lettering against the original, side by side.
  6. Keep the original. Save the small version beside the big one, so you can go back and try another setting.

In a chat tool there is no dial, so the words do the job:

USE WHEN: A CHAT TOOL HAS NO UPSCALE SETTING
Make this exact image larger and sharper. This is a restoration, not a redesign. Keep the same subject, face, pose, colours, lighting and composition. Remove blur and noise only. Do not add objects, text or detail that is not there; leave anything unclear soft.

For 30 more prompts, see the upscale prompts page. In the OpenAI and Gemini pages read for this guide, neither mentions an upscale button; they describe editing and generating images instead.First-hand: on 27 Oct 2026 the Gemini image-generation guide, OpenAI’s image-generation guide and image-prompting guide, and OpenAI’s “Images in ChatGPT” help page were searched for the word “upscal”. None contained it. OpenAI also notes that “Repeated edits can still change details you intended to preserve.”8

What goes wrong when you upscale?

The plain point: the new detail is a guess, so the risks are faces, text and a look that is smoother or busier than the original.

  • Faces change. Topaz turns off its Face Recovery for Redefine creative “in order to maintain the best results from the model”, which tells you the creative mode is not built to keep a face exactly.4 Google’s own research found that people picked an upscaled 16x16 face as the real photo close to half the time, which shows how convincing invented detail can be.9
  • Text breaks. Krea says its Enhancer “cannot reveal what text on a blurry sign actually said”, and a 2025 paper found diffusion-based restoration methods “frequently generate plausible but incorrect text-like patterns”.6
  • Skin and surfaces change. High texture or sharpness can add pores or brushstrokes that were never there (Krea, above). Glossy skin instead? See why AI images look plastic.
  • It is not proof. In 2024 a Washington state judge refused AI-enhanced video as evidence, saying the technology relies on “opaque methods to represent what the AI model ‘thinks’ should be shown”.10 Keep the original and call the enlarged copy enhanced.

What if you want to run it yourself?

The plain point: free, open-source upscalers exist, but you run them yourself and need to be comfortable with code.

Real-ESRGAN describes its aim as “Practical Algorithms for General Image/Video Restoration” and is published under a BSD-3-Clause licence.11 Its ready-made command-line programs have a limit: the project says they may “introduce block inconsistency” because they cut the image into tiles and handle them one by one.11

Questions people ask

What is the best AI image upscaler?

None of the makers’ pages read for this guide says theirs is the best, and no source here ranks them. The useful question is which dial you need: a faithful one for a photo of a person, a creative one for a soft AI picture you want to enrich. Try your image in two tools and compare them at full size; that is this site’s advice.

Can ChatGPT or Gemini upscale an image?

They can redraw a picture from your instructions, but the pages read for this guide describe no upscale button or strength setting. To get a bigger picture directly, Gemini and the OpenAI image API accept larger sizes, as the sizes section above shows. For a faithful result, use the prompt above and compare with the original.

Does upscaling add real detail?

No, it adds predicted detail. Krea puts it plainly: the Enhancer “does not recover what was originally there”.6 See what AI upscaling means.

How many times can you upscale one image?

As few as you can. Krea warns that “Compounding enhancements will compound the AI’s interpretations”, because a second pass works on the first pass, not the original.6

Where to go next

QUICK CHECK · 3 QUESTIONS

Pick an answer to see if you got it. Nothing is recorded.

1. What should you check before you upscale an AI image?

  • Whether the tool can already make a larger picture directly
  • Whether the image is in colour
  • Whether the file name is short
Show the answer

Answer: Whether the tool can already make a larger picture directly. The page says some tools can generate at 2K or 4K, so an upscale may not be needed at all.

2. What choice does almost every upscaler give you, under different names?

  • Faithful or creative
  • Colour or black and white
  • Portrait or landscape
Show the answer

Answer: Faithful or creative. The page says the dial runs from copy my picture to improve my picture, called Subtle or Creative, Resemblance, Creativity or Strength.

3. What should you do after upscaling a picture of a person?

  • Delete the original
  • Upscale it again to be sure
  • Zoom in on faces, hands and text and keep the original
Show the answer

Answer: Zoom in on faces, hands and text and keep the original. The new detail is predicted, so the page says to check faces, hands and lettering at full size and keep the original.

SOURCES
  1. 1Google, Gemini API: image generation, read at source 27 Oct 2026“Gemini 3 image models generate 1K images by default but can also output 2K, 4K, and 512px (05.K) (Gemini 3.1 Flash Image only) images.” Its size table lists a 1:1 4K image for Nano Banana 2.1 as 4096x4096, and says “Gemini 3.1 Flash Lite Image only supports 1K resolution.” ↑a“Gemini 3.1 Flash Lite Image only supports 1K resolution.” (Nano Banana 2 Lite is the name the same page gives this model.) ↑b
  2. 2OpenAI, Image generation guide, read at source 27 Oct 2026“neither edge may exceed 3840 pixels. The total pixel count must be between 655,360 and 8,294,400 (4K). Resolutions above 2560x1440 are experimental.” ↑
  3. 3Midjourney, Upscalers, read at source 27 Oct 2026“With Midjourney V8.2, your SD images start at 1024 x 1024px (using the default aspect ratio). You can then use the Creative and Subtle upscalers to bump up your images to 2048 x 2048px. Note: HD images generate at 2048 x 2048px automatically, and currently cannot be upscaled further.” ↑a“If you choose the Creative upscaler, it will change your image a bit as it adds new details. On the other hand, the Subtle upscaler keeps changes to a minimum, preserving the look of your initial image as closely as possible. Pro Tip: Using the Creative upscaler can sometimes help correct small issues in your image, like awkward hand positions or odd facial expressions.” ↑b
  4. 4Topaz Labs, Generative Models | Topaz Gigapixel, read at source 27 Oct 2026“For use with low quality or AI generated images, use Redefine to add definition and detail. You can prioritize upscaling for either realistic fidelity or creative distinction.” Redefine realistic: “Choose between None or Subtle”; Redefine creative: “Choose between Low, Medium, High, or Max.” ↑a“This model brings the best upscaling fidelity for old and low quality photos.” (Topaz’s own claim about its model.) ↑b“Face Recovery is disabled with Redefine creative in order to maintain the best results from the model”. The second half of this sentence is this site’s reading. ↑c
  5. 5Adobe, Enhance image quality with Generative Upscale in Photoshop, read at source 27 Oct 2026“Select the AI model to power the upscaling. Each model delivers a distinct type of enhancement. Firefly Upscaler: Restores low-resolution images (up to 6144 px × 6144 px) Topaz Gigapixel: Preserves existing detail (available for users with a premium subscription; up to 56 megapixels) Topaz Bloom: Adds new, creative detail (available for users with a premium subscription; up to 9 megapixels)” ↑a“In the Generative Upscale dialog, select the desired output scale (2x or 4x) and adjust parameters as needed.” ↑b
  6. 6Krea, Enhancer (user guide), read at source 27 Oct 2026“Upscale V1 | Fast, neutral upscaler that preserves existing detail well” and “Krea Enhance | Generative upscaler suited for adding detail and improving lower-resolution AI images”. ↑aStrength: “Controls how much additional detail the AI invents. Higher values produce sharper results with more AI interpretation.” Resemblance: “Higher values keep the output closer to the original image. Lower values allow more variation.” Texture (Topaz Generative): “A lower value upscales existing detail; a higher value generates additional texture”. ↑b“Upscaling a painting with high texture will fill in additional brushstrokes that were not in the original”; “adding visible pores and fine lines to skin that was originally clean. The right approach is to start conservatively and increase settings only until you reach the result you want.” ↑c“The upscale factor refers to the increase in width and height, not total file size.” (2x = 4x the size, 4x = 16x); “upscaling in steps -- 2x first, then 4x -- tends to produce more controlled results than jumping directly to 8x.” ↑d“Describe the final image you want to see, not the change you want made.” ↑e“It cannot reveal what text on a blurry sign actually said, for example. Enhancer is a creative tool, not a forensic one.” arXiv, Text-Aware Image Restoration with Diffusion Models, submitted 11 Jun 2025, read at source 13 Oct 2026 (re-used from this site’s AI upscaling page): “Those methods frequently generate plausible but incorrect text-like patterns”. ↑f“it does not recover what was originally there.” ↑g“the second pass operates on the 2x version, not the original. Compounding enhancements will compound the AI’s interpretations, so it is worth being deliberate about sequencing.” ↑h
  7. 7Hugging Face, Diffusers: Super-resolution, read at source 27 Oct 2026“Pipeline for text-guided image super-resolution using Stable Diffusion 2.”; “It is used to enhance the resolution of input images by a factor of 4.”; “A higher noise_level adds more noise to the input, increasing variation but reducing fidelity to the source image.” ↑
  8. 8OpenAI, Image prompting, read at source 13 Oct 2026 (re-used from this site’s AI upscaling page): “Repeated edits can still change details you intended to preserve. Restate those constraints and inspect each result.”↑
  9. 9Google Research, Ho & Saharia, High Fidelity Image Generation Using Diffusion Models, 16 Jul 2021, read at source 13 Oct 2026 (re-used from this site’s AI upscaling page): “We achieve close to 50% confusion rate on the task of 16x16 → 128x128 faces”.↑
  10. 10NBC News, Washington state judge blocks use of AI-enhanced video as evidence in possible first-of-its-kind ruling, 2 Apr 2024, read at source 13 Oct 2026 (re-used from this site’s AI upscaling page): the ruling said the technology relies on “opaque methods to represent what the AI model ‘thinks’ should be shown”.↑
  11. 11GitHub, xinntao/Real-ESRGAN, read at source 27 Oct 2026“Real-ESRGAN aims at developing Practical Algorithms for General Image/Video Restoration.”; listed under “BSD-3-Clause license”; models include “RealESRGAN_x4plus_anime_6B” for anime illustrations. ↑athe portable executables “may introduce block inconsistency (and also generate slightly different results from the PyTorch implementation), because this executable crops the input image into several tiles”. ↑b

Read at source 27 Oct 2026: Midjourney, Upscalers; Topaz Labs, Generative Models | Topaz Gigapixel; Adobe, Generative Upscale in Photoshop; Krea, Enhancer; Google, Gemini API image generation; OpenAI, Image generation; Hugging Face, Diffusers super-resolution; GitHub, Real-ESRGAN.

Re-used from this site’s AI upscaling page (read 13 Oct 2026): Google Research (SR3 faces), arXiv 2506.09993 (text restoration), NBC News (Washington ruling), OpenAI image prompting.

The six-way list, the step-by-step routine, the sample prompt and the advice on which setting to pick are this site’s own, not claims from these sources. No upscaler on this page was tested for this guide, and no tool is ranked.

Free printable AI cheat sheets, and AI news every week.

Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.

SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK
aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE