The Best AI Image Upscalers of 2026–2027: Benchmark With Reproducible Metrics and Prices

Image upscaling is one of the oldest tasks in computer vision. Rare is the spy movie without this scene: an operator marks some area, enlarges it, zooms in and, with one magic button, gets a super-high-res face of the subject. From classical computer vision in the pre-neural-network era to the most advanced multimodal LLMs of today, researchers keep trying to solve this simple-looking and, in reality, almost unsolvable task.

In 2025 there was an unbelievable rise of general-purpose image editing models of all kinds, mostly based on or exploiting multimodal LLMs and diffusion models. The exponential growth in quality started with FLUX.1 Kontext and was marked out by the Nano Banana release. These models can perform any transformation of an image that can be expressed with a text instruction, from the most precise to the most abstract. But however good they were at general tasks, they still lacked precise transformations, preserving the original image and strict prompt following. Especially noticeable were the changes of colours and, most important, of people's identities by the proprietary chatbots (ChatGPT, Gemini and the like). Despite this, back then it looked like all image transformation tasks would be solved in the shortest time.

More than a year has passed, models have improved, preserving the original has become much better, but systematically the same types of problems remain. Image upscaling (super-resolution) is an especially demanding task here: while general editing can often allow some degree of fantasy ("add a cat": which breed and the other parameters are not specified, just a cat), upscaling does not allow any changes of the original. It doesn't matter whether one wants to upscale a photo from a phone for a poster or make a 4K image from an AI-generated picture: usually they don't mean to change anything but the details and the resolution.

For illustration, we asked ChatGPT to make a small photo from a street race four times bigger and to change nothing (the picture). It came back sharp, clean and bright. But look at the runner's number. The real bib says LUT 3029. ChatGPT drew SEINE-MARNE 77, a French county near Paris, and replaced "trucks" on the shirt with another word. None of it is in the photo. Gemini wrote LUT 2022. From the same small copy, one upscaler in our test read all four digits (more tools on this runner).

This is what general AI image editors do as upscalers now. It doesn't matter how exactly you ask and design the prompt: they don't just enlarge the pixels, they try to guess what should be there and paint it in. Often the guess is right. Sometimes it is a different word, a different colour or even a different face. At the same time the results usually look very nice and aligned with the human understanding of aesthetics.

Meanwhile, many specialized image upscalers are on the market. Because the task is historically old, the number of such services is enormous, and it is very hard to find a way through all the models from such different epochs. So we built the most complete upscaler test we know of, with the most popular, successful and precise methods.

The short answer is above. The rest of the article is the precise and reproducible narration, built this way: a core test: 18 tools and settings, the same small, damaged copies of 16 photos, and every result checked against the full-size original that no tool ever saw. Then the full test: three leaders on 135 photos with 567 faces, with every judge on every result and every metric that can be measured on it. Then a set of extremely hard cases: print scans, heavy JPEG, noise, old prints, statues and paintings. Finally, we provide detailed prices at 10, 100 and 1 000 photos a month. The pictures are here ↓

Quick answers.

• What is an AI image upscaler? A tool that makes a picture bigger and has to guess the detail that is not in the pixels.
• What does 4× mean? Four times wider and four times taller: 504 × 336 becomes 2016 × 1344.
• Can a photo be upscaled with nothing changed? Not by any tool here.
• Is there a free one? Upscayl: free and offline, but not for small faces.
• Is it enough for print? 2016 × 1344 prints at about 17 × 11 cm at 300 pixels per inch.
• Need it sharper, not bigger? That is an enhancer, not an upscaler. In MegaPotato the mode is called Refine, and we did not test it here.

The scoreboard: 18 AI image upscalers and settings compared#

Every tool made the same 16 small photos (504 × 336 pixels, lightly damaged) 4× bigger, and every result was checked against the original. The total is out of 100. The formula and every part of it are in How we tested.

The top four are within two points. MegaPotato (82.6), Topaz Wonder 3.5 (81.9), Let's Enhance (81.8) and Crystal (80.8) are a group of their own. The next tool, ChatGPT with a strict prompt (the prompts are in Settings and prompts), is eight points behind (73.2). So we decided to proceed with three of them and make a much wider, harder and more challenging set.

Image quality vs fidelity: aesthetics and preserving the original#

We used mostly real photos in this test. That's not only because it is one of the most important cases, but also because such photos are the hard case. With AI art, a new texture often just looks like more art. With a real photo there is a reference, obviously. So every mistake is easy to see: a new face, a new word, a colour that was not there.

Obviously, ChatGPT has the most aesthetic results, which are, no surprise, among the farthest from the original. Specialized tools are close to the original but less aesthetic (it also depends on the input).

The full test: MegaPotato vs Topaz Wonder 3.5 vs Crystal on 135 photos#

The full test went to three of the four leaders of the core test: MegaPotato, Topaz Wonder 3.5 and Crystal. (Let's Enhance, third in the core test, is web-only, we did not run it on the 135 photos.)

The sample: 135 photos and 567 faces.

• 14 photos from the core test, with the original.
• 16 AI-generated photos (FLUX.1 [schnell]) and 13 photos with no people (animals, nature, buildings), shrunk and damaged the same way as the inputs of the core test (a little blur and noise, saved as a JPEG twice), with the original.
• 29 real low-quality photos as they were taken: old phone photos in low light, early digital cameras, messenger copies, stage and party shots. No high-resolution original exists.
• 19 old prints and film scans with dust, fading and grain, made smaller for the test. No clean original exists.
• 44 natively small photos fed exactly as they were posted: 37 Flickr photos from 2001–2007 and 7 frames from 2005–2009 web videos. No high-resolution original exists.

Across them: people (103 photos), crowds and groups (30), faces (567 faces: 348 of them 10–24 pixels wide in the input, 173 of 25–59 pixels, 46 larger), signs and text (20), night, stage and low light (24), old photos (21), AI images (16), animals (20), nature (26), cities, buildings and rooms (55) and web-video frames (7). We left out still lifes (products and food): this sample is about people, places, animals and nature.

Where there is no clean high-resolution original, the judges compared each result with the input, and only the measures that need no original were used: closeness to the original (LPIPS, A-FINE, same-person score) comes from the 43 photos that have one. Each measure is averaged over the photos it can be measured on, and the total is the weighted sum of these averages. The n column says how many photos or faces each line covers.

Results by kind of photo. The same formula, run on each part of the sample on its own. Where a kind has no faces or no originals, those parts are left out and the other weights are scaled up to add to 100 %, so compare the tools within a kind, not one kind with another:

What the numbers say:

• MegaPotato leads the full test, by 3 points. It is especially good on real low-quality photos (67.2 against 62.8 for Wonder and 46.1 for Crystal, 29 photos), on night and stage light (70.1 against 64.8 and 55.2, 24 photos), on AI images, on text and signs, and on crowds.
• Wonder 3.5 is level with it on old prints (74.0 against 73.9, 19 photos) and on the natively small photos (75.4 against 75.5, 44 photos), and ahead on web-video frames (67.6 against 65.5, 7 frames). Crystal leads only on photos with no people (74.2, 13 photos), and falls far behind on real low-quality photos.
• The blind Opus 5.5 judge agrees with the order: MegaPotato 7.13, Wonder 3.5 6.99, Crystal 5.97 of 10 (135 photos), and so does the Qwen3.8 judge on zoomed areas (7.24, 7.21, 6.83).
• Faces: the same split as in the core test. Wonder 3.5 and Crystal stay closest to the person in the original, as measured by an ensemble of face-recognition models. MegaPotato's faces got the best scores from the Qwen3.8 face judge and were the least often called a different person.
• A-FINE gives MegaPotato its biggest lead, because it finds MegaPotato's results more natural. A-FINE (CVPR 2025)^[1] judges two things, how close a result is to the original and how natural it looks, and when a result looks more natural than the original it gives naturalness a much bigger weight. Taken apart on the 43 photos of the full test that have an original, A-FINE found MegaPotato's result more natural than the original photo on all photos.

How we tested the AI upscalers#

Basically, all test photos fall into two categories:

1. High-res, high-quality photos that we degraded with different methods (JPEG, blur, noise and so on) and resized to a low resolution.
2. Natively low-res photos.

Zoomed in:

The core test (16 photos, 18 tools and settings).

• The same damaged input for everyone: 504 × 336 pixels, a little blur and noise, saved as a JPEG twice, like a photo that has been shared a few times. The AI image was only shrunk.
• The same job for everyone: 4 times bigger (2016 × 1344). Topaz, Crystal, Let's Enhance, MegaPotato, Krea, Upscayl and Magnific were set to 4×. Recraft has no size setting but returned exactly 4×. ChatGPT and Gemini can't be told a size, so we scaled their results to 2016 × 1344.
• Default settings, one run per photo. Topaz: six models. ChatGPT: a simple and a strict prompt. We did not change any default parameters to keep the test clean.

The full test (135 photos, 3 leaders). The photos are described above. Every result was made 4× bigger in one run: MegaPotato in its web app, Wonder 3.5 and Crystal through fal.ai with their default settings.

The judges, both blind. Using strong multimodal AI models as graders^[2] is an established method in AI research^[3]. We used two different ones, so that one model's taste can't decide the result.

• Claude Opus 5.5 (max effort)^[4] saw every result under a random letter, as a whole picture and zoomed areas, next to the original (or the input, where no original exists), and gave a score from 0 to 10. One result per photo was shown twice to check its consistency.
• Qwen3.8^[5] (27B) scored zoomed areas and every face next to the original, on the same 0–10 scale, and answered for every face: is it still the same person?

The measurements. Every face was found once on the original with SCRFD^[6] (on the enlarged input where there is no original) and cut out the same way from every result, so a face a tool lost can't escape the test.

• Same person? The average similarity from four face-recognition models (three on most faces): ArcFace^[7] (trained on Glint360K^[8]), MagFace^[9], AdaFace^[10] (trained on WebFace12M^[11]) and AuraFace^[12] (1 = the same face, different people score about 0–0.3).
• Face quality: eDifFIQA^[13] (the fast successor of DifFIQA^[14]), which scores how clean and recognisable a face is, without the original.
• Closeness to the original: LPIPS^[15] and A-FINE^[1] in the total; DISTS^[16], DreamSim^[17], DINOv2^[18], TOPIQ^[19] and the CIEDE2000 colour difference^[20] as checks.
• Does it look good? (no original needed): MUSIQ^[21] and NIQE^[22] in the total, MANIQA^[23], CLIP-IQA+^[24] and TOPIQ as checks.

The metrics ran with the open IQA-PyTorch library^[25] and the models' own code.

The total (0–100). 28 % the Opus score, and 8 % each for: the Qwen3.8 score on zoomed areas, the Qwen3.8 score on faces, the same-person score, the share of faces Qwen3.8 called a different person, face quality, LPIPS, A-FINE, MUSIQ and NIQE. The judges' scores are multiplied by 10. Every other measure is scaled so that the best of the 18 tools and settings of the core test gets 100 and the worst 0. The full test uses the same scale, so its totals mean the same thing. A value better than the best of the core test counts as 100, a value worse than the worst as 0. The weights are our choice. Every part is shown, so you can weigh them your own way. We fixed the weights after the core test and before the full test. Opus got the most weight because it is an extremely capable proprietary model, and on the core test its scores matched our own eyes best.

About reproducibility. Before scoring, every result was lined up with the input: a result shifted by more than 0.75 pixels was moved back (Wonder 3.5's results are often shifted by about 0.8 pixels). The numbers repeat: see How to repeat this test.

Before and after: detailed examples#

The crops below are from the core test unless the figure says otherwise.

Faces: which AI upscaler kept the person's identity#

Under every result, in small type: Opus 5.5 judge, the blind judge's score (0–10) for the whole result the crop comes from. face sim., how close that face is to the person in the original as the face-recognition models measure it (0–1, different people score about 0–0.3), where there is a face. MUSIQ, the looks-good score of the whole result (0–100).

On a close portrait almost every tool did a good job: the image carries much more information, so the task is easier.

Small faces, on the other hand, are extremely hard, so the tools split into groups very clearly. In a crowd or an old group photo a face is only 10–35 pixels wide, so the model mostly has to guess.

What the face numbers say. The core test has 67 faces on 9 photos, 10–290 pixels wide in the copy the tools got. For each one we measured the same-person score, the Qwen3.8 face judge's score, and whether the judge called it a different person:

• Face recognition and the AI judge disagree in an interesting way. A blurry plain resize gets a high "same person" score, because blur keeps the rough shapes. But the judge discounts such cases. A sharp face has to be guessed, and every guess moves it a little away from the original. That is a trade-off no model can outplay.
• Face quality can't tell the leaders apart. eDifFIQA scored the faces of Crystal, Wonder 3.5, Let's Enhance, MegaPotato and ChatGPT at 0.56–0.59, the same as the real faces (0.58). Distorted or blurry faces scored much lower (Upscayl 0.39, plain resize 0.27). But that is more of a looks metric: it rates how clean and recognisable a face is, not whether it is the same person.

Creative models are a different story. They don't try to keep the person's identity at all in most cases:

Text and numbers: the most confident mistakes#

Small text is where upscalers are the most dangerous. The result looks readable, so you believe it.

Big text is easier than small, obviously:

Colours: who kept the colours of the photo#

Animal fur and AI-generated images#

Invented details: what the tools added that was not in the photo#

Every generative tool invented some detail that never existed in the original. General-purpose models did it more often, specialized ones more rarely, but none is immune.

• MegaPotato: made-up logos and characters where the text is too small to read.
• ChatGPT: new objects and decorations (curtains, pictures, company names) and a different sky and light.
• Topaz Wonder 3.5: turns statues, dolls and painted faces into living people and photographs. Wrong digits where the text is too small.
• Topaz Standard V2, Low Resolution V2 and Standard MAX: tend to remove accessories from small faces.
• Topaz Bloom 2: new clothes and objects on people, statues and tables.
• Recraft Creative: a new scene: new people, plants, windows and menus.
• Magnific Creative: a new scene too: new people and text, a bust turned into an urn.

Extremely hard cases: old prints, heavy JPEG, noise (57 pictures)#

The full test measures the average. So we also made a few tests with extreme degradation, to see how the models behave in the worst case: 57 pictures, made small (288 × 216 or less) and damaged on purpose, or just hard by nature. Print scans with halftone dots, heavy JPEG, low-light noise, old prints with dust and scratches, textured paper, tiny faces, sunglasses, statues, dolls, paintings and comics. The originals are public-domain or free-licence photos (Library of Congress, Wikimedia Commons, The Metropolitan Museum of Art, Pexels). Only the three leaders ran here, and blind AI judges picked the best result for every picture. For the pictures shown below the Opus 5.5 judge also scored every result in a separate blind pass, with the core test's rubric; its score and MUSIQ are under the panels. The changed faces you will see are the tools' output, not the people. We show them only to document what the tools do.

Badly damaged faces are our weak spot: on 8 such pictures MegaPotato changed the person on 6 (a new age, a new face or opened eyes). On very small inputs nobody kept Einstein's face, and MegaPotato also left a band of streaked pixels along two edges of every result.

In short: Wonder 3.5 and MegaPotato split the hard cases exactly evenly, 31 first places each of 57, and Crystal took 7. Print scans, heavy JPEG and noisy photos: start with Wonder 3.5. Old prints, statues, dolls and textured paper: MegaPotato. Paintings: Crystal, it keeps the brushwork. With damaged faces nobody was safe.

Which AI upscaler should you use, by kind of photo#

• Real low-quality photos as they were taken (old phone photos, early digital cameras, party and stage shots): MegaPotato, and by a wide margin (67.2 against 62.8 for Wonder 3.5 and 46.1 for Crystal, 29 photos).
• Night, stage and low light: MegaPotato (70.1 against 64.8 and 55.2, 24 photos).
• Shared, compressed or small photos of people: MegaPotato for the sharpest, most natural faces. Topaz Wonder 3.5 or Crystal when the face must stay as close to the original as possible (higher same-person scores).
• Old faded prints and old family photos: Wonder 3.5 or MegaPotato: level on 19 old prints (74.0 and 73.9). MegaPotato gave the sharpest faces in the hard cases but sometimes turns heavy damage into patterns. No tool reliably removes dust and scratches.
• Scans of newspapers and magazines: Topaz Wonder 3.5 or MegaPotato.
• Statues, dolls, paintings and drawings: MegaPotato keeps stone eyes as stone. For paintings, Crystal keeps the brushwork. Wonder sometimes turns them into "living" ones.
• Landscapes, animals and buildings with no people: all three are close (Crystal 74.2, MegaPotato 72.8, Wonder 3.5 72.6).
• Photos with names, numbers or signs: don't trust any tool with small text. A word or number that suddenly became readable is a guess, not a recovery. Small text is what we are working on for the next version of MegaPotato.
• A big print from a small file: at 300 pixels per inch, 2016 × 1344 pixels prints at about 17 × 11 cm (6.7 × 4.5 in). MegaPotato goes to 8×, Let's Enhance to 16×, Topaz Gigapixel to 6×.
• Thousands of images through an API: MegaPotato ($0.091 a photo at 1 000 a month), Topaz ($0.26 through its own API from $50 a month, or $0.24 through fal.ai) or Crystal ($0.35) for photos. Recraft Crisp ($0.033) only for clean graphics. We did not test speed or reliability at volume.

AI upscaler prices: what they really cost#

Prices for one photo hide two things: what you must buy first, and whether you can use the tool without writing code. So here is the cost of a whole month of the small job of the core test (a 504 × 336 photo made 4× bigger, 2.7 MP), which every tool can do. Prices are from each vendor's own website or API as of 8 October 2026 (fal.ai's Topaz prices, Topaz's app plans and Clarity AI's plans re-checked on 10 October, Topaz's API plans on 13 October, Krea's plans and MegaPotato's own prices on 14 October), before tax. The pages are in the sources.

What this means in practice. If you can write code, MegaPotato's API is the cheapest of the careful tools: $0.104 a photo (the average of three sizes at 100 photos a month, no batch discount through the API; the web app has it, and there the same photos cost $0.083), against $0.54 for Topaz Wonder 3.5 through Topaz's own API ($0.24 through fal.ai) and $0.35 for Crystal's API. Without code the choice is narrower. Crystal's website starts at $13 a month billed yearly ($19 monthly). Wonder 3.5 for photos exists only through an API, fal.ai pay as you go or Topaz's own from $50 a month (Topaz's own apps have Wonder 3, which we did not test). So MegaPotato and Let's Enhance are basically the only two top tools you pay for once, in a browser, no subscription, and MegaPotato is cheaper at every size we priced. For really heavy use, around 1 000 photos a month, Topaz's flat plans with Wonder 3 are hard to beat (Gigapixel $149 a year, Topaz for Web $199 a year list, $109 on sale in October): Topaz calls the renders unlimited and publishes no monthly cap, only a 32 MP export limit and two images at a time. Whether that holds at thousands of renders a month we could not test. At 100 photos a month MegaPotato is already cheaper.

Every AI upscaler we tested, one by one#

Short notes on each tool: what it is, what it did well, where it broke. The scores are in the scoreboard and the prices in the price tables, no point repeating them here.

1. MegaPotato AI image upscaler: first on both tests and the cheapest of the leaders#

MegaPotato is our own tool: a web app and an API, one model that does 2×, 4×, 8× or Refine at the same size, chip packs instead of a subscription.

It is strongest on real low-quality photos, night shots and AI images, the faces come out natural, and it cleans old prints (paper grain, dust) better than the others. The weak side is small faces - they drift a bit more from the real person than with Crystal and Wonder 3.5, in dark damaged photos it changed ages and sometimes faces, and on very small inputs it leaves streaked edges.

2. Topaz Wonder 3.5: the most faithful faces, but only through an API#

Topaz Labs is probably the best-known name in upscaling (part of Adobe since September 2026), with three families of models that behave very differently: Precision, Generative and Creative. Wonder 3.5 is the one that matters, and it is API-only, Topaz's own API or fal.ai, where we ran it. The apps have Wonder 3.

The most faithful faces of all, the only tool that read the runner's number, the best on halftone scans and the most careful with damaged faces. The problems: it treats statues and dolls as people and paintings as photographs, changes people's ages, and leaves old faded prints soft. The other Topaz models are a different story: Standard V2, Low Resolution V2 and Standard MAX turn small faces to plastic and tend to remove accessories, High Fidelity V3 barely touches a small photo (it is made for big ones), Bloom 2 redraws faces.

3. Let's Enhance: almost as faithful, and easy to use#

Let's Enhance is one of the oldest web upscalers, web only (its API is sold separately as Claid, we did not test it), with a handful of modes; we used Prime, the default. You pay with credits, as packs or a plan.

Careful and faithful: the face judge liked its faces most of all, and it invented nothing big. Where it loses is sharpness: its results are the softest of the four leaders, both to the zoom judge and by the looks-good measures, and the faces stay a bit further from the real person than Wonder's and Crystal's.

4. Crystal Upscaler (Clarity AI): very faithful on normal photos, weak on noise and print#

Crystal is Clarity AI's closed model, tuned for faces and products: an API paid per image, and a website on a plan.

The closest to the originals by the image metrics - it keeps the person and the colours, adds nothing, and is the best when there are no people in the frame. But noise and JPEG blocks it keeps, only sharper, halftone dots stay dots, and in our hard cases it made dark sunglasses see-through and children in old photos look like adults. Not for scans, damaged prints or noisy photos.

5. ChatGPT as an upscaler: great at first sight, but it changes important details#

ChatGPT has no upscale button: you ask for a bigger version and the image model draws a new picture.

The best-looking results of all at first sight. But the well-known issue with redrawing is still there - it shifts colours, makes people younger and writes text with full confidence, and a strict "don't invent anything" prompt does not stop it. Fine for a nice picture, not for family photos, documents, products or anything with names and numbers.

6. Gemini app (Google Nano Banana): careful with faces, small files#

Gemini is not an upscaler either, it redraws the whole picture.

The most careful of the chat models: it kept faces and colours. The files are small and soft, though, and it writes clean, wrong numbers and made-up words.

7. Recraft (Crisp and Creative): cheap for graphics, not for small faces#

Recraft is a design tool with two upscalers: Crisp sharpens what is there, Creative redraws. You can't choose the output size.

Crisp is fine on AI art and clean edges, and it is the cheapest API in the test, but it distorts small faces. Creative simply draws another photo, new people and nonsense words included. Logos, graphics and AI art in bulk, not people.

8. Upscayl: free and offline, but not for small faces#

Upscayl is a free, open-source desktop app that runs Real-ESRGAN-type models on your own graphics card: no account, no watermark, the photos never leave your computer (three of the bundled models are marked non-commercial).

Free, private, decent on AI art and clean edges, and it invents no words. Small faces come out distorted and small text stays blurred. Free offline upscaling of clean images, that's it.

9. Krea: a big toolbox, we tested its own enhancer#

Krea is a creative platform; its Enhancer has Krea's own models and partner ones (Topaz, Magnific). We tested Krea's own.

A lot of tools in one place, if you already generate there. On photos its own Enhance drifted far from the real faces, drew black lines across the 1914 faces and made red chalk grey; it ended near the bottom of our test.

10. Magnific: the famous one draws a new picture. Its photo mode over-sharpens#

Magnific is the upscaler everyone names first, two modes, Creative and Precision. We ran both through ComfyUI's Magnific nodes at the nodes' defaults; the node returned 2× whatever we asked, so our 4× is two 2× passes.

Creative makes a rich new picture, good for concept art and illustrations. For a photo that is exactly the problem: it draws a new one, with other people and other words. Precision keeps the scene but piles sharp invented texture on everything. With the defaults neither came near the top; Precision at lower detail settings may do better than our run.

How to repeat this test#

What you need to run the test again is on this page and in four small files:

• The photos: each of the 137 photos of the two main tests, with its source, licence and how its small input was made, in photos.csv. How each kind of input was made is in the appendix. The hard cases are described in their section.
• The tools: where each tool ran, with its settings and prompts, in Settings, prompts and lining up.
• The scoring: the judges, the measurements and the formula, in How we tested. Every tool's numbers for both tests are in the full test's table and the appendix.
• Every score: each tool on each photo in scores.csv, each face in faces.csv. README.txt says what every column is and how they add up to the tables on this page. The data is free to use under CC BY 4.0, with a link to this page.

Version 1.0, 14 October 2026. The tools ran in early October 2026. Prices are as of 8 to 14 October 2026, each source carries its date. Changes to the numbers will be listed here.

We checked that the numbers repeat. Scoring the saved results of the core test a second time through the same code gave every number within 0.15 % of the first run, photo by photo. Asked again about 201 faces with the same prompts and crops, the Qwen3.8 judge gave the identical score 85 % of the time, and the same "same person?" answer for 99.5 % of the faces. Claude Opus 5.5 saw one result per photo twice under two letters. In the core test its two scores were identical on 9 of the 16 photos and half a point apart on the others. The upscalers themselves can give a different result on a second run.

Appendix: metrics, settings, prompts and photo sources#