Steam Capsule Art to Spec: Dimensions, Legibility, Rejections
Steam's capsule art has fixed dimensions and a legibility bar most AI key art fails at 120x45 — the size that actually shows up in a player's library.
Steam never actually reviews your capsule art at the size you upload it. It reviews the size it auto-generates from that upload — a 120×45 thumbnail, derived by downscaling your 462×174 small capsule, that ends up carrying a game's entire visual identity in a library grid, a recommendation shelf, or a search results row. That's 15,540 pixels doing the job a poster does at full size, and it's where most AI-generated capsule art quietly fails: not rejected outright, just unreadable the moment it actually ships.
The exact dimensions, and why the small one matters most
Per Steam's own asset documentation, the required set is fixed and non-negotiable:
| Asset | Dimensions | Where it appears |
|---|---|---|
| Header capsule | 920×430 | Top of your store page, recommendations |
| Small capsule | 462×174 | Library, search, widgets — auto-generates 120×45 and 184×69 |
| Main capsule | 1232×706 | Homepage carousel, featured placements |
| Vertical capsule | 748×896 | Sale pages, seasonal events |
| Screenshots | Minimum 5, at 1920×1080, 16:9 | Store page gallery |
The line that matters most is the parenthetical: Steam doesn't let you upload a 120×45 image directly. It derives that thumbnail from your 462×174 small capsule by downscaling. Whatever legibility problem exists in your source art doesn't get fixed at the smaller size — it gets amplified. Art that reads perfectly fine at 462×174 on a monitor while you're designing it can turn into an illegible smear at 120×45 the moment a real player scrolls their library.
What actually breaks at 120×45
Run any capsule art through a mental downscale to 120×45 and the failure points are predictable: fine detail under roughly 14 effective pixels tall vanishes completely, busy backgrounds compete with the logotype for the same handful of surviving pixels, thin linework smears into gray mush, and color contrast that looked crisp at full resolution collapses once anti-aliasing averages neighboring pixels toward each other.
AI-generated capsule art is unusually prone to this failure specifically because generation models are optimized to reward the opposite of what a thumbnail needs. Rich detail, complex composition, and layered depth read as "high quality" at full size and in most model training objectives — and are exactly the properties that disappear first under a hard downscale. A capsule that impresses you at 1232×706 with intricate background elements and soft gradients is frequently the one that turns into a gray rectangle at 120×45.
Text rendering: the other place generation breaks
Not every image model can render a legible logotype, and this matters more for capsule art than almost any other AI image use case, because the title text often is the thumbnail's entire job at small sizes. General-purpose photorealistic models routinely mangle letterforms — warped strokes, merged characters, invented glyphs that resemble letters without being letters.
Versely's catalog carries three models specifically built for this: Recraft 4.1 Text to Image, Recraft V4, with its explicit typography support, and Ideogram V4, built for accurate text rendering alongside posters and logo work. If your capsule needs the game's logotype baked directly into the generated art, start with one of these rather than a general photoreal model — it's a different job than illustrating a hero subject, and treating it as the same job is where a lot of AI-assisted capsule art goes wrong before it ever reaches the downscale problem.
A design approach that actually survives the shrink
The fix is procedural, not stylistic: test at 120×45 before you finalize, not after.
- Design in reverse. Generate or compose your key art at whatever working resolution you like, but before calling any version final, downscale a crop to both 120×45 and 184×69 and judge it there — not at the size you're staring at in your editor.
- One bold subject beats a busy scene. A single, high-contrast silhouette survives a hard downscale; a layered composition with multiple focal points doesn't.
- Separate the illustration from the logotype. Rather than trusting one generation to nail both a compelling hero image and crisp title text simultaneously, treat them as two layers: generate the hero art with a strong single subject and a controlled palette, then add the logotype separately — either from a text-capable model's dedicated pass or from an actual font set as vector type. Two components you can each verify independently beats one generation you're hoping does both jobs at once.
- Keep the palette limited. A wide color range looks rich at full size and turns to noise at 120×45; two or three dominant tones with strong contrast between subject and background hold up far better.
Screenshots are a different problem — and AI generation is the wrong tool
The five required screenshots follow a completely different rule, and it's worth being precise about it because the mistake here isn't a legibility issue, it's a policy violation. Steam's documentation is explicit: screenshots "should exclusively show the gameplay of your game," specifically ruling out concept art, pre-rendered cinematic stills, and images carrying marketing copy or award badges. Menu screens are allowed only when they're integral to how the game is played.
That means capsule art and header art can reasonably be AI-assisted key art, but the five required screenshots cannot be — they need to be real captures from the actual, running game. It's a distinction solo and small-team developers blur more often than you'd expect, generating a striking concept render and submitting it where an actual gameplay capture belongs. Keep the two workflows separate: AI-assisted illustration for the capsule set, real captures for the screenshot gallery.
A Versely walkthrough: building the capsule set as one consistent system
- Generate the hero composition first, at your widest ratio. Start with the main capsule's 1232×706 (roughly 1.75:1) using a text-capable model like Ideogram V4 or Recraft V4 if the logotype needs to live inside the art itself. Favor one bold subject and a tight palette from the first prompt — you're designing for the smallest derivative size from the start, not the largest.
- Derive the rest of the set from that same composition rather than regenerating each size independently. Crop or re-frame the header (920×430), small capsule (462×174), and vertical capsule (748×896) from the hero art so the set reads as one consistent visual system — Steam displays several of these side by side, and mismatched styles across your own capsule set are as damaging as any single weak asset.
- Test the small capsule at its derived sizes before finalizing. Shrink a copy of your 462×174 crop down to 120×45 and 184×69 and actually look at it there. If the logotype or subject disappears, fix the source art — don't ship it and hope.
- Use Versely's text-to-image tool to iterate variants quickly, and lean on the best AI image generator comparison if your first model choice isn't holding up at small sizes — legibility at 120×45 is a legitimate axis to pick a model on, not an afterthought.
FAQ
What is the actual size Steam capsule art gets judged at?
Functionally, 120×45 — the thumbnail Steam auto-generates by downscaling your 462×174 small capsule. It's the size shown across library grids, search, and most in-store surfaces, and it's the size your art needs to survive, not the 462×174 you upload.
Can I upload a 120×45 image directly to control exactly how it looks?
No. Steam derives the 120×45 and 184×69 thumbnails automatically from the 462×174 small capsule you provide — there's no separate upload for the smallest sizes, so all the design work has to happen at the source resolution with the downscale in mind.
Which AI models render legible text for game logotypes?
In Versely's catalog, Recraft 4.1 Text to Image, Recraft V4, and Ideogram V4 are built for accurate text and typography rendering, unlike general-purpose photoreal models that frequently distort letterforms. Pick one of these when the logotype needs to be part of the generated art itself.
Can I use AI-generated images for my Steam store screenshots?
No — Steam's screenshot requirements call for real captures of actual gameplay, explicitly excluding concept art and pre-rendered cinematics. AI generation is a reasonable tool for capsule and header art, but the five required screenshots need to come from the running game.
Why does capsule art that looks great in my editor fail once it's live?
Because you're judging it at full resolution while players mostly see it downscaled to 120×45. Rich detail, soft gradients, and busy compositions read as high quality at full size and disappear first under a hard downscale — the fix is testing your art at the smallest derived size before finalizing it, not after.
Design for the thumbnail first and the poster will take care of itself — build your capsule set with the AI image generator, then shrink it down before you ever hit submit.