Gallery
The Gallery will show finished AI pictures, clips and home-lab photos, each with the prompt, the model, the machine and the license. Albums open later. This page shows how to write the card now.
Albums
No pictures yet. The first albums open with membership.
Beginner
One picture and the words that made it
A gallery card is one finished picture with the words that produced it written beside it. Those words are the prompt: the instruction you typed into the image tool. Under the picture, keep four lines. Name the tool (the website or app). Name the model (the specific system inside that tool, copied from the tool's own label). Paste the full prompt, including any line you marked as things to avoid. Add the license you grant other people for this picture, or all rights reserved if you grant none.
When the four lines are filled in, save them as a text file next to the picture. If the tool shows a seed, copy that number too. A seed is the starting number for the tool's random choices. The same prompt, the same model, and the same seed are what let someone else aim at the same result.
You can practice before any account exists. Hugging Face Spaces runs image, video, and speech apps in the browser. For pictures you may study and reuse, open each file page in the Wikimedia Commons category for artificial-intelligence art and read that file's license. Files on Commons carry different licenses, so the category page is only the index. Poly Haven HDRIs are lighting environments for 3D scenes, released as CC0, which means you may use the environment files in commercial work without asking. The preview pictures and the site logo on Poly Haven stay under the site's own terms. Save the card as a text file next to the image.
Write the card for one image
I made one image with an AI tool. Ask me, one question at a time, for: the tool name, the model name as shown in the tool, the full prompt, any line I used for things to avoid, any seed, the image size in pixels, and the license I want to grant other people. Then output a plain-text gallery card with those fields, a one-sentence description of the picture, and a single block I can paste back into the same tool. If I do not know a field, write UNKNOWN. Do not invent a model name, a seed, or a license.
Intermediate
Before and after, five boxes, and an agent map
At the intermediate level, a card shows each step in order. Place the first output and the last output side by side and name the single change between them: a rewritten prompt, a new seed, a different model, or a longer run. A run is one press of the generate button. Write the settings the tool actually used, in the tool's own names: steps (how many refinement passes), sampler or scheduler (the rule for each pass), and image size in pixels.
Lay the work out as five boxes, left to right. Box 1 is the brief in one sentence, written by you. Box 2 is the full prompt. Box 3 is the settings and the seed. Box 4 is the output file name, the time it took, and what you rejected. Box 5 is the license for the output and the license of every source picture you fed in. If you fed in a photo of a person, write whether you have their permission to publish that face.
An agent flow is a job split across several model calls. Draw it as boxes and arrows: the request, the planner that breaks the request into steps, each tool the planner calls (search, an image model, a code runner), the checker that accepts or sends the step back, and the file you keep. Put the model name on every box. Save the diagram next to the five-box card. Another person should be able to redraw your arrows from the text alone, without opening your account.
Turn one job into five boxes
I will paste notes from one AI image or video job, including any failed attempts. Turn them into five labeled boxes: 1 brief, 2 full prompt, 3 settings and seed, 4 output file, time, and what I rejected, 5 licenses for the output and for every source file. Then add a before/after note that names the single change between the first output and the last. If the job used several model calls, also list boxes and arrows: request, planner, each tool, checker, final file, with a model name on every box. Mark any missing fact UNKNOWN. Do not invent settings, seeds, times, or permissions.
Expert
Video, cost, and the machine on the desk
An expert card treats the machine and the clock as part of the work. For a video, list each shot: duration in seconds, frames per second, model, resolution, and wall-clock time from the first click to the finished file. Add paid software, cloud rental, and electricity if you metered it. Leave a field blank when you did not measure it.
For a text or speech model, report tokens per second. A token is a chunk of text the model reads or writes, often a short word or part of a word. Record prompt length, output length, batch size (how many requests at once), quantization, and the software build. Quantization stores the model's numbers with fewer bits so the file fits in memory. Write the bit width and the method name from the model card, such as 4-bit or 8-bit, and link the card. A model card is the model's readme: intended use, training data, evaluation, and license. Record the model license and the license on your picture as two lines.
A home-lab photograph belongs in the gallery when the caption carries the spec: graphics-card name, motherboard slot or PCIe adapter, cooler, operating system and version, and driver or runtime version. If the frame shows another person's face or belongings, get permission before you publish. Where the tool can write them, keep Content Credentials, the C2PA record of origin and later edits, and the IPTC fields for the AI prompt, the prompt writer, and the AI system. IPTC is the photo-metadata standard that stores those facts inside the image file.
Build a video and lab spec from my notes
Here are rough notes from a video render or a home AI computer. Produce a spec sheet with one row per shot (duration, frames per second, model, resolution, wall-clock time) and a cost block limited to paid software, cloud rental, and metered electricity. Add a tokens-per-second block only if my notes include a text or speech model: prompt length, output length, batch size, quantization bit width and method, software build, and a link slot for the model card. Separate the model license from the output license. Add a hardware block: graphics card, slot or PCIe adapter, cooler, operating system and version, driver or runtime. Use UNKNOWN for anything I did not measure. Do not estimate watts, money, or tokens per second. End with a rights log: people visible in frame, permission status, and whether Content Credentials or IPTC AI fields were written into the file.
Spec card to copy
Use these rows on every card. Write UNKNOWN when the tool hides a value.
| Field | What to write |
|---|---|
| Model | Name and version from the tool, plus a link to the model card when one exists |
| Hardware | Graphics card, or browser and laptop if that is the whole machine |
| Quantization | Bit width and method, or none |
| Prompt | Full text, and any avoid-line the tool stores separately |
| Time | Wall-clock for one image, or for each second of video |
| Seed | The number, or tool hid it |
| License | The license you grant on this file, and the model license on its own line |
| Sources | Links and licenses of pictures, HDRIs, or audio you fed in |
Rights log
Keep a short log with the file, even when you publish only the picture.
- Date, tool, and the name you want public. Leave the account email off the card.
- Prompt writer, if someone else wrote the words.
- Every person whose face or voice is in the frame, and whether you have permission to publish.
- License identifier for the output, using a Creative Commons deed or all rights reserved, and a link to the deed.
- Model license, copied from the model card, including a non-commercial or share-alike limit if the card states one.
- A training-data claim only when the model card itself states it.
- Whether the file contains a C2PA Content Credential or the IPTC AI prompt fields (prompt text, prompt writer, AI system, AI system version).
What an album will ask for
When an album opens, a submission is one card with the spec table filled in, the original file, and the rights log. A blank measured field stays on the card when it says UNKNOWN. A seed, a speed, or a license goes on the card when the tool or the model card showed it. Ratings, once the control exists, score whether the next person can repeat the work from the card: the prompt, the settings, the machine, and the license.
If Polaris buys the shared station, long video renders that do not fit a home PC could be compared there. About the shared station
Sources
- Wikimedia Commons: Artificial intelligence artPublished AI images and videos. Open each file page for author and license. Search all AI images on Commons. Each file has its own license.
- Wikimedia Commons: Reusing content outside WikimediaHow to credit a Commons file, when share-alike applies, and which non-copyright limits can still block a reuse. The Foundation does not warrant the license label.
- Poly Haven HDRIsLighting environments for 3D scenes. The asset files are CC0, confirmed on the Poly Haven license page. Preview renders, avatars, and logos on the site are excluded from that grant.
- Hugging Face SpacesImage, video, speech, and other AI apps you can run in the browser while you learn what to write on a card.
- Hugging Face model cardsWhat a model card contains: intended use, datasets, evaluation, and the license field you copy onto the spec card.
- Creative Commons licensesThe six public licenses, from attribution-only to non-commercial with no derivatives, with links to the legal deeds.
- IPTC Photo Metadata StandardFields stored inside a photo file. Version 2025.1 adds AI prompt information, AI prompt writer, AI system used, and AI system version used.
- C2PA Content CredentialsThe open standard for a content credential: a record of origin and later edits that travels with the file. The specification is linked from this page.