We Made a Complete Discord Server Art Kit with MAI-Image-2.6 for Under a Dollar

The first draft of our server banner, generated with MAI-Image-2.6. It has a problem that we fix later in the article.

A Discord server is a community, but it is also a brand. The banner is the first thing new members see, the icon sits in hundreds of server lists, role badges tell people who is who, and custom emoji become the community's private language. Larger servers commission artists for all of this. Most smaller ones use a stock icon, a free banner template and the default emoji, and look like every other server.

We wanted to know whether a current image model could close that gap. So we invented a gaming community, the Night Owls, and built its full visual kit with MAI-Image-2.6: a banner, an icon, four role badges, an event poster, a welcome card, a rules card, a nine-emoji sheet and a mascot with seasonal variants. We compared the full model with the cheaper Flash version where it made sense, checked every word of text, and recorded what each request cost.

The whole kit took 17 requests and cost $0.89. Every one of the 26 text items we asked for was spelled correctly. One request failed in a way you should know about before you try this yourself.

The tools and the method

MAI-Image-2.6 comes in two versions with the same parameters: mai-image-2.6 and the faster, cheaper mai-image-2.6-flash. Both generate one image per request at 1K or 2K, accept aspect ratios from 1:4 to 4:1, and can take up to five reference images for edits. We generated everything at 2K and downscaled afterwards, because Discord shows banners, icons and emoji at small sizes and a large source gives you room to crop.

The API is asynchronous: you submit a request, get a task ID and poll for the finished image. In our run, MAI-Image-2.6 took 29 to 63 seconds per image, usually around 45 to 50, and Flash took 21 to 29 seconds. At 2K the full model was billed between $0.064 and $0.067 per image and Flash $0.033. Edits that used a reference image were cheaper: $0.044 on the full model.

A quick note on Discord's formats before we start, because they shape every prompt. The server banner is a wide 16:9 image that Discord shows small and partly covered by the server name, so its edges and lower area should be calm. The server icon is shown as a circle, so anything important must sit well inside the frame. Emoji are tiny, so each one needs a single bold idea and a thick outline.

Test 1: the server banner

We began with the banner, using the same prompt on both models.

Discord server banner for a gaming community, wide format. The exact text "NIGHT OWLS" in bold glowing letters on the left, and a stylised owl with glowing amber eyes perched on a neon-lit city rooftop at night on the right. Purple and teal palette, sleek esports style.

Test 1: both banners spelled the name correctly, and both added slogans we never asked for.

Both banners looked like something a professional esports team might use. The owl had glowing amber eyes on both, the city was rendered in purple and teal neon, and "NIGHT OWLS" was spelled correctly in a heavy display style. The full model gave the letters a cleaner, more technical outline; Flash used a rougher brush style that some communities would prefer.

But both models filled the empty space with text of their own. The full model added "GAME LATE. STAY SHARP. TOGETHER.", three small labels with icons, a line at the bottom reading "TOGETHER WE OWN THE NIGHT", and a vertical sign with Japanese characters. Flash added "PLAY TOGETHER. STAY AHEAD. BELONG." and three more labels. The invented text was spelled correctly, but on a Discord banner it is a real problem: it competes with the server name and makes the banner busy at small sizes. We fix this in Test 9.

Test 2: the server icon

The icon is the most important single asset of a server, because it appears everywhere. It has to read at 48 pixels and survive being cropped to a circle.

Discord server icon: a bold, simple owl head emblem with glowing amber eyes in purple and teal, centred inside a circle with generous padding so it survives a round crop, flat vector style, dark background, no text.

Test 2: both icons were centred with enough padding to survive Discord's round crop.

Both icons followed the brief precisely: a bold, symmetrical owl head, no text, a circle with padding all round, and a dark background. The full model used teal for the ear tufts and white and grey for the face, which gave strong contrast at small sizes. Flash used more purple and a glowing ring, which looked richer at full size but slightly muddier when we scaled it down to 48 pixels. Either would be a good starting point; both cost less than seven cents.

Both icons at real Discord sizes, and the emoji sheet from Test 5 sliced into nine 128-pixel emoji.

Test 3: role badges

Role badges need small, legible labels, which is where image models traditionally fail.

A set of four Discord role badges in a row on a dark background, each a rounded shield with an owl motif and a label: exactly "ADMIN" in gold, "MODERATOR" in teal, "VIP" in purple and "MEMBER" in grey. Clean flat vector game UI style.

Test 3: four badges, four colours, four correct labels.

This was a flawless result. All four badges used the requested colours, the owl motif was consistent across the set, and every label was spelled correctly, including the long "MODERATOR", which fitted inside its shield without being squeezed. The set is ready to be cut into four separate images.

Test 4: an event poster and a welcome card

Next, two assets with information that must be exactly right: an event announcement with a time and a channel name, and a welcome card with a member number.

Gaming event poster for a Discord community. Exact text: "FRIDAY RAID NIGHT" as the title, "9 PM UTC" below it, and "Voice channel: #war-room" at the bottom. Dramatic fantasy art of a party of four adventurers facing a dragon, purple and teal neon accents.

Test 4: the time, the channel name with its hash sign and the member number with its comma were all correct.

Both were correct in every detail. The poster rendered "FRIDAY RAID NIGHT" in a brushed neon style, "9 PM UTC" between two small diamonds, and "Voice channel: #war-room" in a pill-shaped label, with the hash sign and hyphen intact. The fantasy art, four silhouetted adventurers facing a purple dragon, was dramatic and fitted the palette. The welcome card showed "WELCOME TO THE NIGHT OWLS" and "You are member #1,024" with the comma in the right place, plus a friendly version of the owl wearing a scarf.

The welcome card also shows one of the more interesting behaviours of the model: it invented a cute, cartoon owl here, quite different from the menacing owl on the banner, because the prompt said "friendly style". If you want one mascot across every asset, you need a reference image, which is what Test 6 does.

Test 5: an emoji sheet

Custom emoji are what make a server feel like a community. We asked for nine expressions of one character on a single sheet.

An emoji sheet of nine cute owl stickers in a 3 by 3 grid on a plain white background, the same purple owl character in each with a different expression: happy, laughing, crying, angry, sleepy, surprised, thumbs up, heart eyes, thinking. Consistent thick outline sticker style, no text.

Test 5: nine expressions, one consistent character, in a single image.

All nine expressions were present and easy to tell apart, including the subtle ones: the sleepy owl has closed eyes and a "z", the thinking owl has a question mark and a raised wing, and the angry owl has crossed wings and an anger symbol. The character is the same in every cell, with the same colours, the same white sticker border and the same proportions. Generating the set in one image is the key: it forces the model to keep one design, which is much harder across nine separate requests.

To use them, slice the sheet into nine squares and scale each to 128 pixels. At that size the thick outlines and simple shapes still read well, which is exactly why we asked for that style. An even three-by-three split worked for eight of the nine; one cell picked up the edge of a neighbouring sticker, which a manual crop fixes in seconds.

We generated every asset through the MAI-Image-2.6 endpoint, which serves both the full and Flash models with the same parameters, so switching between them for drafts and finals is a one-word change. Each completed task returns its billed cost, which is how we built the cost table later in this article.

Test 6: a mascot and its seasonal variants

Many servers have a mascot that changes for events: a Halloween costume, a winter look, a trophy after a tournament. We created a character sheet first, then used it as the reference for three edits.

The same owl mascot from the reference jumping with joy and holding a golden trophy that reads exactly "CHAMPIONS", same style, plain white background.

Test 6: the character stayed recognisable in all three variants, but the winter edit dropped the headset.

The character sheet itself was clean and useful: front, side and back views of a purple owl wearing a teal gaming headset, with consistent proportions across the three views. The Halloween and trophy edits kept everything that makes the mascot recognisable, including the headset with its small owl logo, and added exactly what we asked for. The trophy was engraved "CHAMPIONS", correctly spelled, on both the full model and on Flash.

The winter edit was charming but not faithful: the model replaced the headset with the red bobble hat instead of putting the hat on top of it. That is a small change, but a mascot's headset is part of its identity, and a community will notice. If an accessory matters, name it in every edit prompt, for example "keep the teal headset". Each edit cost $0.044 on the full model and $0.019 on Flash.

Test 7: the transparency trap

Stickers and emoji usually need a transparent background, so we asked for one directly.

A purple owl mascot wearing a teal headset, cartoon style, isolated on a transparent background, PNG with alpha channel.

Test 7: the "transparent" background is a painted checkerboard, not real transparency.

This is the one request in the kit that failed completely. The image looks transparent in a thumbnail, because the background is the familiar grey and white checkerboard. But it is painted: the file is an ordinary RGB image with no alpha channel, so the checkerboard would appear on Discord exactly as you see it here. This is a common behaviour of image models, which learned from millions of images that show transparency as a checkerboard. The fix is simple: ask for a plain white or flat-colour background, as we did for the emoji sheet, and remove it with a background-removal tool.

Test 8: a rules card

A rules card is a block of text with a little decoration, which makes it a good stress test for text accuracy. We asked for a title and five exact lines on both models.

Discord server rules card on a dark purple background with the exact title "SERVER RULES" and five numbered lines exactly as follows: "1. Be kind to everyone", "2. No spam or self-promotion", "3. Keep it safe for work", "4. Use the right channels", "5. Have fun". Clean modern layout with a small owl icon.

Test 8: both models rendered all six text items correctly, including the hyphen in "self-promotion".

Both cards were correct, title and all five rules, including the hyphen in "self-promotion" and the numbering. The layouts differed: the full model put each rule in its own rounded row, like a settings panel, while Flash used a centred card with the key words highlighted in lighter purple. Flash's card was arguably the more attractive of the two, and it took 22 seconds and cost $0.032, against 42 seconds and $0.065 for the full model. For text-heavy assets like this, Flash was the better value in our test.

Test 9: fixing the banner

Back to the problem from Test 1. Rather than regenerating the banner and losing the owl and the city we liked, we passed it back to the model as a reference with a single instruction.

Recreate this Discord banner with only the words "NIGHT OWLS" and no other text anywhere, keep the owl, the city and the colours.

Test 9: one edit removed every invented slogan and kept the composition.

The edit removed every slogan, label and sign and kept the composition, the owl, the rooftop and the neon palette almost exactly. The name itself was redrawn and is still spelled correctly. The result is a banner that works at Discord's small banner size, with the name clearly readable and nothing else competing with it. It cost $0.044 and took 36 seconds. The general lesson applies to every asset in this article: if a model adds text you did not want, an edit with the words "no other text" is usually faster than starting again.

Full model or Flash?

We ran four of the assets on both versions of the model, with identical prompts, so we can compare them directly.

The pattern is consistent with what we have seen in other tests of this model. Flash costs about half as much and is usually about twice as fast, and on text accuracy it was just as good: every requested word was correct on both. The full model's advantage shows up in fine visual detail and contrast, which matters most for the one asset that has to work at a tiny size, the icon. A sensible workflow is to draft everything on Flash, choose the designs you like, and regenerate only the icon and the banner on the full model.

Speed matters more than it might seem for a project like this. Designing a community's look is iterative: you try a colour, change your mind, try another style. At 20 to 30 seconds per image, Flash makes that loop feel interactive, and at three cents per image, nobody needs to worry about how many attempts they make.

The kit and what it cost

The complete kit: 17 images for $0.89.

The walkthrough video

Walkthrough: building a Discord server art kit with MAI-Image-2.6 (about 3 minutes).

Practical tips for your own server

1. Say "no other text". Both models added slogans to the banner when the prompt left space for them. The instruction removed them in one edit.
2. Generate emoji as one sheet. A single image with nine cells kept the character consistent in a way separate requests would not.
3. Use a character sheet as the reference for every variant. And name the accessories you want to keep in each prompt.
4. Never ask for transparency. Ask for a flat background and remove it afterwards.
5. Use Flash for text-heavy cards. It spelled the rules card perfectly at half the price and twice the speed.
6. Design for the crop. Ask for padding on icons and calm edges on banners, because Discord will crop and cover both.
7. Check small sizes. Scale icons and emoji down to their real size before you upload them; detail that looks great at 2K can disappear at 48 pixels.

The bottom line

For less than a dollar, MAI-Image-2.6 produced a Discord kit that would have cost a small community a meaningful amount to commission: a banner, an icon, badges, posters, cards, nine emoji and a mascot with variants. Every word we asked for was spelled correctly, the character stayed consistent where we gave it a reference, and the one serious failure, the fake transparency, has a simple workaround. The model's main habit to manage is its enthusiasm for adding text you did not ask for. Tell it not to, and edit when it does anyway. If you want to build a kit for your own server, or compare MAI-Image-2.6 with other image models through a single account and API key, you can find them all on APIMart.ai.