Image creation
Image tools turn a written description into a picture, and some of them can change a picture you already have. This page shows how to run one prompt on two tools, how to score hands, lettering and color, and where public boards publish head-to-head votes.
Beginner
Same prompt, two tools
A prompt is the text you give an image tool. The model is the program that turns that text into pixels. A pixel is one colored dot in the file. Two models given the same prompt often return different pictures, so the first test is to paste one prompt into two tools and save both files with the prompt beside them.
Check five things in the same order every time. Count the fingers on every visible hand, including hands wrapped around a cup or partly out of frame. Read every word in the picture, including shop signs, book spines and labels on clothing, and match the spelling to the prompt. Check each color you named: if the prompt asks for a red mug, the mug in the file has to be red. Count the objects: three cups means three cups. Note anything the tool added on its own, such as extra people, watermarks or logos.
Also write down the tool name, the date, and the width and height in pixels. Many tools default to a square such as 1024 by 1024. If you want a poster, ask for a 16:9 frame, sixteen units wide for every nine of height.
Style words steer the look. Photograph, watercolor, flat icon and technical diagram are four different requests. Use one style word per test so you can see what that word did.
The market and the job board open later. You can run this two-tool check in any image chat today.
Two-tool picture check
Make one image from the description below. If you cannot create images, return the description unchanged so I can paste it into an image tool, then list the five checks I should score: finger count, spelling of every word, named colors, object count, and anything I did not ask for. Description: A photograph of one left hand holding a red ceramic mug by the handle. Five fingers are visible. Steam rises from the mug. The background is plain gray. No text, no logo, no other people.
Intermediate
A fixed battery, then an edit
A prompt battery is a fixed list you reuse on every model. Changing the words between runs tests the sentence as well as the model. Ten prompts are enough if each one targets a common miss: readable text, a hand on a named object, a reflection, an exact count, or two close colors side by side.
Resolution is the width and height of the file in pixels. Ask every model for the same size, and record the size it actually saved. Compare sharpness only when those numbers match. A wide 16:9 frame and a square are different compositions.
Editing starts from a picture you already have. Write what must change and what must stay. Use two scores. Change: did the sign get the new words? Keep: did the face, background and framing stay? A new person in the frame fails Keep.
Public boards work by preference votes. Arena's text-to-image board shows two pictures from one prompt and asks a person to pick one. Those picks become the score. A Text Rendering view covers prompts that demand words. The image-edit board scores a change to one uploaded picture, and another tab scores edits that use several pictures. Artificial Analysis runs a separate text-to-image board on pairwise votes and prints a score with a range. Check the date. New models move the order.
If the model file can be downloaded, it has open weights. Read the license next to the file. Apache 2.0 and MIT allow broad reuse. A non-commercial license blocks client work.
Edit scorecard
I will test an image edit. If you can edit images, apply only the change below once I upload a picture. If you cannot, give me a blank score sheet I can fill in by hand. Change: the shop sign must read exactly FRESH BREAD DAILY. Keep: the same building, the same people, the same camera angle and the same time of day. Score Change from 0 to 2: 0 means the words are wrong, 1 means they are partly right, 2 means every character matches. Score Keep from 0 to 2: 0 means the scene was replaced, 1 means small unwanted changes, 2 means the scene stayed. Then list every character that differs from FRESH BREAD DAILY.
Expert
Vision, letters, adapters, rights
A vision test asks a model about a picture that already exists. A generation test asks it to make the picture.
MMMU is a public set of college questions with pictures. The project site describes 11.5K items, six disciplines, 30 subjects and 30 picture types, including charts, diagrams, sheet music and chemical structures. The CVPR 2024 paper says students collected the items from exams and textbooks. Use MMMU to measure reading a diagram. Use a prompt battery to measure drawing one.
For text inside a picture, write the exact words you asked for and count the characters that appear in the right order. Arena's Text Rendering category gathers public votes on pictures that were supposed to contain words.
LoRA means Low-Rank Adaptation: a small set of extra weights trained while the base image model stays frozen. In the Diffusers library, only those new weights are updated, and the adapter file is much smaller than the full model. Rank is the width of the extra matrices. Higher rank can fit more of a style and uses more memory. The training guide is marked experimental.
In the United States, the Copyright Office stated in January 2025 that an AI output is protected only where a human determined sufficient expressive elements. A human work that remains visible can qualify, and so can a human's creative arrangement or edit of the output. The Office says that merely supplying prompts does not provide that authorship. Other countries write different rules. On Wikimedia Commons, open the category Artificial intelligence art and read the license on each file page before you reuse the file.
Adapter test plan
Write a test plan I can run before and after I attach a LoRA to an image model. A LoRA is a small set of extra trained weights; the base model stays frozen. The plan must include the base model name, the adapter file name, the rank if I know it, one prompt that should show the new style, one prompt that should stay ordinary, the pixel size I will lock for both, and a five-line score for hands, spelling, colors, object count and unwanted changes. If you can generate images, make both pictures at that same pixel size and fill in the score. If you cannot, output only the blank plan.
Ten prompts worth repeating
Use the same ten prompts, the same pixel size and the same style word on every model. Save the prompt text with the file. Prompt 10 is an edit of the picture you got from prompt 1, so score Change and Keep separately.
| # | Prompt to paste | What to score |
|---|---|---|
| 1 | A photograph of one left hand holding a blue ceramic mug. Five fingers are visible. Steam rises. Plain gray background. No text. | Finger count, mug color, extra objects |
| 2 | A shop window. The only text anywhere in the picture is a sign that reads exactly: FRESH BREAD DAILY. | Every character, in order |
| 3 | Top-down photograph of three red apples and two green apples on a wooden table. No other fruit. | Count and color of each apple |
| 4 | Side view of one cyclist in a yellow jacket. Both wheels are fully visible. The wet street shows a reflection of the bicycle. | Wheels, jacket color, reflection |
| 5 | Close photograph of one person, both ears visible, no jewellery, neutral expression, plain background. | Ear count, extra accessories |
| 6 | A flat icon of a house, a tree and a sun. Solid colors only. No gradients. No text. | Style, object count, stray detail |
| 7 | A technical line diagram of a bicycle on a white background. Black arrows labeled exactly FRAME, WHEEL, SADDLE and HANDLEBAR. | Spelling of the four labels |
| 8 | Night street with one street lamp. The lamp light is warm yellow. The sky is deep blue. No moon, no other lights, no text. | Color of lamp and sky, extra lights |
| 9 | One paperback standing upright. The cover title is exactly NORTH STAR. The author line is exactly A. BERG. No other words. | Both strings, character by character |
| 10 | Edit the picture from prompt 1. Move the mug handle to the left side. Keep the same hand, the same blue mug, the same gray background and the same framing. | Change (handle side) and Keep (everything else) |
How a result row will be published
Published comparison rows could start if Polaris buys a shared test machine. Each row will name the model, the prompt number, the pixel size, the seed if the tool prints one, the date and the machine name. A seed is the starting number that lets the same random draw be repeated. Until those rows exist, run the battery above in any image tool you can open yourself.
For an edit, fill two boxes before you judge taste. Change asks whether the requested difference is in the file. Keep asks whether the parts you named as fixed are still the same picture. A replacement of the whole scene scores zero on Keep.
A shared DGX Station would run this fixed prompt battery, and LoRA fine-tunes, on open image models that need more graphics memory than a laptop has, and publish each run with the date and the machine name. About the shared station
Sources
- Arena text-to-image leaderboardBlind preference votes on pictures made from text, with a separate Text Rendering view for words inside images. Ranks move, so read the date on the page.
- Arena image-edit leaderboardPreference votes for edits of one uploaded picture, plus a multi-image edit view. Useful when you are scoring Change against Keep.
- Artificial Analysis text-to-image leaderboardA second public board that ranks text-to-image models from pairwise human votes and shows a score with a range.
- MMMU benchmarkCollege-level questions with charts, diagrams and other pictures, for testing whether a model can read an image. The site links the dataset and the paper.
- MMMU paper (PDF)The CVPR 2024 paper: how the 11.5K questions were collected and what the test is meant to measure.
- Diffusers LoRA training guideOfficial guide to Low-Rank Adaptation for image models: small extra weights, rank, and the warning that the training interface is experimental.
- U.S. Copyright Office, Part 2 announcementJanuary 2025 statement of United States copyright law: an AI output is protected only where a human determined sufficient expressive elements, and merely supplying prompts does not provide that authorship.
- Wikimedia Commons: Artificial intelligence artA public category of AI-related art files. Each file page states that file's own license. Do not assume one license covers the whole category.