A good upscale prompt names what must stay the same first (face, pose, composition, colours), then what to clean up (blur, noise, compression artifacts), and tells the tool not to invent anything. Upload your image to any tool that accepts one, paste this, and send:
How to do it, and what to type: this page is what to type. For the steps, the settings and which upscaler to pick, read how to upscale an AI image or how to upscale AI video.
The prompt follows the same order OpenAI gives for editing a picture: OpenAI’s image prompting guide says to “identify what should change and what must stay the same” and to “list the details to preserve, such as identity, geometry, layout, lighting, or labels.”1
If it changes too much, use the stricter “change nothing, just sharper” prompt. Below it are prompts for renders, photos of a screen and portrait sizes, plus 24 restoration and style prompts: 30 in all, counting the universal prompt once (it appears again lower down in comma-list form with a negative line). One more, for old drawings and sketches, is in the questions at the end. They are starting points: none carries a recorded test date. In Midjourney you do not need a prompt: you press an upscale button, and Midjourney says the Subtle upscaler keeps changes to a minimum, preserving the look of your initial image as closely as possible.2
Upscaling can also transform an image: the same img2img process can restore an old photo, recover a video still or shift a generated image into a new look. Results vary by model.
- Change nothing, just sharper the strict prompt for chat tools with no denoising slider; compare the result with the original.
- Renders and CGI cleans jagged edges and render noise, and keeps it a render.
- Photos of a screen removes moire lines, the pixel grid and glare, and straightens the angle.
- Portrait sizes for a tall post or a vertical story, extending only the background.
- Universal prompt with a negative line the same universal prompt as a comma list, for tools that read a negative line.
- Photo restoration old family photos, video stills and historical archive images.
- Cinematic film film looks described by light, lens and colour rather than a filmmaker's name.
- Editorial and fashion magazine fashion, minimal lifestyle and sports editorial looks.
- Mood and atmosphere neon, golden hour, blue hour, noir and misty forest moods.
- Product and commercial jewelry, food and architecture shots.
- Artistic and painterly oil painting, watercolor and vintage analog looks.
- Portraits and faces beauty-campaign skin detail, documentary portraits and corporate headshots.
- Old drawings and sketches keeps the line work and the paper, and removes only dust, scan noise and blur.
- ChatGPT, Gemini and Midjourney upload and paste a prompt in chat tools; in Midjourney, press an upscale button instead.
- Stacking styles combine elements from several prompts and the model blends the looks.
What Upscaling Actually Does (and What It Can Do Beyond Resolution)
Before you upscale anything: if an image reads as fake, enlarging it makes the wrongness bigger rather than smaller — that is a lighting and surface problem, not a resolution one. What photorealistic, hyperrealistic and cinematic actually mean covers the distinction, including why "hyperrealistic" in an upscale prompt pushes away from realism rather than toward it.
Most people think of AI upscaling as resolution improvement — making a small or blurry image larger and sharper. That's the baseline use case. But the img2img upscale workflow is actually a controlled image transformation that can simultaneously enhance resolution, improve quality, and shift the aesthetic direction of an image.
At denoising 0.35–0.50, the model preserves composition and identity while improving texture detail. At 0.50–0.65, it starts meaningfully shifting the visual style. These are working ranges, not measured thresholds, and the point where style starts to shift varies by model. Understanding this range is what unlocks the full range of prompts below.
This works especially well for:
- Old blurry family photos — restore clarity, recover detail, improve skin and background quality
- Video stills and screenshots — extract a frame from any video and upscale it to usable image quality
- AI-generated images — take any generated image and add a final quality pass before delivery
- Low-res historical images — recover archival photography with modern rendering quality
- Product photography — sharpen, relight, and elevate existing product shots
Upscaling can transform, not just enlarge. At denoising 0.35–0.50 it keeps composition and identity; at 0.50–0.65 it starts shifting the style. It will not fix an image that already looks fake.
Universal Parameters (Apply to All Prompts Below)
1 — The NEGATIVE: line on each prompt below only applies to models that read one. Of the tools listed above, Midjourney does (via --no). FLUX.2 does not — Black Forest Labs states it "does not support negative prompts" — and ChatGPT Image has no negative field either. On those, fold the exclusion into the positive line: not "no artificial sharpening" but "natural edge transitions, no halo ringing described as smooth gradient falloff." See what the vendors actually say.
2 — Denoising, steps, CFG and sampler apply to img2img pipelines you control — ComfyUI, A1111, a hosted API with those parameters exposed. In a chat interface like ChatGPT Image there is no denoising slider; the equivalent lever is how strongly you tell it to preserve the original. The values below are a starting point for the former, not a setting you can type into the latter.
3An earlier version of the box above recommended the DPM++ 2M Karras sampler “for all upscale work”. No source supported that, so it now says to keep your pipeline’s default. The other values are working starting points, not measured settings.
The settings only apply where you control them. NEGATIVE lines work only on models that read one, like Midjourney, and chat tools such as ChatGPT Image have no denoising slider.
Faithful Upscale Prompts: Change Nothing, Just Sharper
The plain point: chat tools such as ChatGPT and Gemini do not have a denoising slider, so the only way to stop them changing your picture is to say so, in detail. ChatGPT, for example, works from your words: OpenAI’s help page says you can upload an existing image and describe the changes you want ChatGPT to make.4 Treat the result as a careful redraw rather than a pixel-for-pixel enlargement, and compare it side by side with the original.
Want to change one thing, such as the outfit, and keep everything else? That is an edit, not an upscale: see image edit prompts that change only one thing.
Upscale prompt for a render or CGI image
A 3D render, game capture or CGI still needs different clean-up from a photo: jagged edges and render noise rather than film grain. This prompt keeps it a render. If you want the render to look like a photograph instead, that is a transformation: use the cinematic prompts further down.
Upscale prompt for a photo of a screen
A phone photo of a monitor or TV carries problems a normal photo does not: wavy moiré lines, the screen’s pixel grid, glare and a tilted angle. Name each one.
Upscale prompts for portrait sizes (4:5 and 9:16)
Say the size in the prompt and also set it in the tool. ChatGPT has an aspect ratio picker: OpenAI’s help page says to use the aspect ratio picker to choose a different aspect ratio, or include your desired aspect ratio in your prompt.5 In Midjourney’s Edit model the output follows the first image you upload unless you add --ar.6
The universal prompt with a negative line
The same universal prompt as at the top of this page, in the comma-list form many image apps expect, with a negative line for tools that read one (Midjourney reads it through --no).
Name what must not change. In a chat tool, the prompt is the only control you have, so list face, pose, background, composition and colours, and say what to leave soft rather than invent.
Photo Restoration Prompts
Three restoration starting points. For a full set of old photo restoration prompts, tool by tool, including how to stop a face changing, see the dedicated page.
Cinematic Film Prompts
These prompts describe the light, lens and colour rather than naming a filmmaker; what a director’s name does to a prompt explains why.
Editorial & Fashion Prompts
Mood & Atmosphere Prompts
Product & Commercial Prompts
Artistic & Painterly Prompts
Portrait and Face Prompts
How to Use These Prompts in ChatGPT, Gemini (Nano Banana) and Midjourney
The plain point: in ChatGPT and Gemini you upload the image and paste a prompt; in Midjourney you press an upscale button instead. Each tool documents its own way in.
ChatGPT (ChatGPT Images)
Upload the image in a chat, paste the universal or the faithful prompt, and send. To work on one area only, open the image and use the selection tool, then describe the change; OpenAI warns that the highlight can spill over.5 Developers reach the same models through OpenAI’s image API.7
Gemini (Nano Banana)
In the Gemini app, choose Create images from the tools menu, upload your picture and paste the prompt.8 Through the Gemini API you can also ask for a larger output size (1K, 2K or 4K, written with a capital K), which matters more than typing “8K” into the prompt.9
Midjourney
For an image you made in Midjourney, open it and press Subtle (keeps the look) or Creative (adds new detail and changes the image a little). Both double the size.2 In V8.2 that takes a standard square image from 1024 × 1024 to 2048 × 2048 pixels, and HD images already start at 2048 × 2048 and cannot be upscaled further.2 For a photo from outside Midjourney, upload it to the Edit model and use the faithful prompt as a written instruction.6
Prompt in chat tools, button in Midjourney. ChatGPT and Gemini redraw from your words, so the faithful prompt matters there; Midjourney’s Subtle upscaler is its “change nothing” option.
Questions people ask
What is the best prompt to upscale an image?
No single prompt suits every tool, but a reliable starting point is one that lists what must stay the same (face, pose, composition, colours) before asking for sharper detail and less noise: the universal prompt at the top of this page. In Midjourney you skip the prompt and press Subtle, which Midjourney says keeps changes to a minimum.2
What prompt enhances image quality?
Ask for the fixes by name: “remove blur, noise and compression artifacts, restore fine detail in skin, hair and fabric”, and say what to keep. Words like “8K” describe a look; the real output size is set by the tool, such as the 1K, 2K or 4K sizes Google documents for Nano Banana in the Gemini API.9
How do I upscale an image without changing it?
Use the “change nothing, just sharper” prompt in a chat tool, then compare the result with the original, because the tool redraws rather than enlarges. In Midjourney, choose Subtle rather than Creative: Midjourney says the Creative upscaler will change your image a bit as it adds new details.2
Does writing “8K” in a prompt make an 8K image?
No. The prompt changes the style; the pixel size comes from the tool’s own settings. Midjourney V8.2, for example, tops out at 2048 × 2048 pixels for a square upscaled or HD image.2
Can AI restore an old photo?
It can sharpen, clean and recolour one, but a face can come back looking like a different person, so keep the original and compare. The restoration prompts above are a start; old photo restoration prompts has the full set by tool. This is the site’s own advice, not a vendor claim.
What is a good AI prompt for upscaling an old drawing or sketch?
Ask for the lines to stay exactly where they are and for the clean-up to stop at dust, scan noise and blur. A drawing has no face or skin to protect, so the thing to protect is the artist’s line work and the look of the paper. List those details to keep, as OpenAI’s guide advises for any edit, and say what to leave faint rather than guess:
Compare the result with the original at full size: a tool that redraws can straighten or invent lines. This prompt is the site’s own, not a vendor example.
Combining Prompts: Stacking Styles
You can combine elements from multiple prompts above. The model will blend the aesthetics proportionally. Examples that work well together:
- Cinematic Film + Portrait: motivated natural light + Ultra-Detailed Skin = cinematic portrait with natural skin quality
- Golden Hour + Product: Golden Hour + Luxury Product = warm commercial product shot with ambient appeal
- Noir + Documentary: Noir + Raw Truth = authentic documentary noir portrait
- Cyberpunk + Editorial Fashion: Cyberpunk Neon + Vogue Editorial = editorial fashion in dystopian environment
You can stack styles. Combine elements from several prompts and the model blends the looks, such as the Cinematic Film prompt’s motivated natural light with Ultra-Detailed Skin for a cinematic portrait.
Related on this site: Why Do AI Images Look Plastic? · Inpainting vs Outpainting · How to Upscale an AI Image
- 1OpenAI, Image prompting guide, read at source 3 Nov 2026“For edits, identify what should change and what must stay the same.” and, under prompting fundamentals, “list the details to preserve, such as identity, geometry, layout, lighting, or labels.” The wording of the prompt itself is this site’s own. ↑
- 2Midjourney, Upscalers, re-read directly at source 13 Oct 2026“the Subtle upscaler keeps changes to a minimum, preserving the look of your initial image as closely as possible.” ↑a“Both will double the size of your original image. If you choose the Creative upscaler, it will change your image a bit as it adds new details.” Steps: “Start by clicking on any image you’ve made on your Create or Organize page.” ↑b“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.” and “HD images generate at 2048 x 2048px automatically, and currently cannot be upscaled further.” ↑c“the Subtle upscaler keeps changes to a minimum, preserving the look of your initial image as closely as possible.” ↑d“Both will double the size of your original image. If you choose the Creative upscaler, it will change your image a bit as it adds new details.” Steps: “Start by clicking on any image you’ve made on your Create or Organize page.” ↑e“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.” and “HD images generate at 2048 x 2048px automatically, and currently cannot be upscaled further.” ↑f
- 3Vendor documentation, re-read at source 11 Sep 2026. Black Forest Labs, FLUX.2 prompting guide: "FLUX.2 does not support negative prompts. Focus on describing what you want, not what you don't want." OpenAI's image prompting guide names no negative-prompt field for any GPT Image model and instead tells you to "State exclusions such as unwanted text, logos, or watermarks" in the prompt itself. Midjourney, No, re-read directly at source 13 Oct 2026“The --no parameter is a tool you can use to tell Midjourney what you don’t want in your image.” Until 12 Oct 2026 this note said the page had not been read. CORRECTED 11 Sep 2026 — the model list above said "Nano Banana Pro 2"; Google's name for the Gemini 3 Pro Image model is Nano Banana Pro. ↑
- 4OpenAI Help Center, Images in ChatGPT, read in a browser 12 Oct 2026 and again by direct fetch 13 Oct 2026, when the page loaded without a browser“You can also upload an existing image and describe the changes you want ChatGPT to make.” ↑
- 5OpenAI Help Center, Images in ChatGPT, read in a browser 12 Oct 2026“Use the aspect ratio picker to choose a different aspect ratio, or include your desired aspect ratio in your prompt.” ↑a“To edit an image, select an image generated by ChatGPT Images or upload an existing image, then describe the change you want.” and “Highlights are not always precise, and edits may extend beyond the area you selected.” ↑b
- 6Midjourney, Edit Model, re-read directly at source 13 Oct 2026“The Edit model will automatically try to match the aspect ratio of the first image you upload. To change this, simply include your desired aspect ratio in the prompt using the --ar parameter.” ↑athe Edit model lets you “Make changes to an existing image using written instructions.” and is “compatible with Midjourney versions 8.1 and 8.2.” ↑b
- 7OpenAI, Image generation (API guide), read at source 12 Oct 2026“The API lets you generate and edit images from text prompts using gpt-image-2.5-sunburst and gpt-image-2.5-flare.” ↑
- 8Google, Nano Banana 2 - Gemini AI image generator & photo editor, read at source 12 Oct 2026“To access Nano Banana, select ‘🍌Create images’ from the tools menu.” “Then add a prompt or upload an image to edit.” ↑
- 9Google, Nano Banana image generation (Gemini API), read at source 12 Oct 2026“High-resolution output: Built-in generation capabilities for 1K, 2K, and 4K visuals.” and “You must use an uppercase ‘K’ (e.g. 512px (05.K), 1K, 2K, 4K).” ↑a“High-resolution output: Built-in generation capabilities for 1K, 2K, and 4K visuals.” and “You must use an uppercase ‘K’ (e.g. 512px (05.K), 1K, 2K, 4K).” ↑b
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