Midjourney for Game Key Art and Steam Headers: A Full Workflow

Making a game's key visual (KV) and Steam header art with Midjourney comes down to a clean division of labor: Midjourney V7 generates the main visual and mood base art, locking in the art direction in one pass, while a version with text gets finished off by a model that renders text reliably. KV art runs at a 16:9 landscape ratio, generating 4 images per batch to pick a composition from; the Steam header targets the roughly 460x215 landscape ratio, matched using 16:9 or wide-format prompts; local flaws get boxed off and fixed separately; the game title and tagline get added last. I ran this whole workflow on Flux Art — an all-in-one AI visual generation workbench that aggregates 50+ leading global image and video models under one account. Midjourney V7 is directly and stably accessible there, with output up to 4K, no watermark, and commercial use allowed. The division of labor is simple too: Midjourney V7 handles the main visual and mood base art, Nano Banana 2 handles local inpainting for flawed spots, and GPT Image 2 — reliable at rendering text — finishes the text version that needs the game logo and title text overlaid.

I've worked in indie game publishing for eight years — from the early days moving overseas launch assets for studios, to now running a small team that signs indie titles and manages storefront launches and UA creative. Promotional KVs and Steam store header art are the most communication-heavy part of a launch: art draws a concept, an outside artist finalizes it, then it gets revised again to pass platform review — a single key visual can drag on for half a month. Since bringing Midjourney into the early concept stage over the past couple of years, the efficiency of that phase has changed completely, though I've also hit plenty of style-mismatch problems along the way. This post lays out the workflow I've settled on, plus a real failure case and how I fixed it.

What exactly are a game's promotional KV and Steam header, and what do they require?

Let's separate the two asset types first — they're both core storefront assets, but the requirements differ.

The KV, or Key Visual, is the game's main promotional image — a single image that has to convey the world, the protagonist, the mood, and the art style all at once. It shows up on the official site homepage, the main art area of the Steam store page, social media promo headers, and convention booth visuals — it's a player's first impression of the game. The KV's job is to "set the tone" — the composition needs a clear focal point, the mood needs to grab attention, and there needs to be negative space left for text to sit on top later. That's why it leans toward a wide landscape frame, with 16:9 being the most versatile ratio, one that can be cropped into several other sizes afterward.

The Steam header is a different matter. A game's Steam store page has a whole set of images: the large banner at the top of the store page, the capsule image in the library, the small thumbnail in search listings — the ratios all differ, but the main header art is close to that squat 460x215 landscape ratio. It's more "sales-driven" than the KV — players scan past it in a fraction of a second while scrolling through store listings, so the image has to communicate the game's name, genre, and tone instantly. That's why Steam headers almost always carry the game logo and title text, and the composition has to leave a clean spot for that text to land.

The shared hard requirement for both types is that the art style has to match the actual game seamlessly. What a player sees in the store has to match what they actually experience in-game — pairing a pixel-art game with a photorealistic 3D header, and having players discover the mismatch after clicking in, is a direct path to refunds and bad reviews. This is exactly where Midjourney is most likely to trip up: it's widely recognized for strong stylistic range and artistic expression, but without anchoring the art direction, it will happily wander off into a different style on its own.

Midjourney V7 is widely regarded as a stylized model — artistic flair and creative expression are its calling card, giving it a real edge for mood base art and concept visuals where "feel" matters most. But it has a well-known, publicly documented weakness: text inside the image tends to come out wrong. Ask it to spell a game's title directly into the artwork, and it's likely to misspell or garble it most of the time. That's exactly why the text version needs to be split off and handed to a different model — this is a pattern that shows up repeatedly in hands-on testing, not an isolated case.

Tools are no longer the scarce resource. According to CNNIC's 57th Statistical Report on China's Internet Development, the number of generative AI users in China reached 602 million by December 2025, up 141.7% from December 2024. Now that anyone can open an image-generation page, what separates publishers isn't "knowing how to use AI" — it's whether they can anchor the style accurately and run a smooth workflow that produces art that both fits the game and is ready to ship.

Worth spelling out too are the pain points of the traditional approach. The old path for a single KV went: art writes a concept brief, an outside artist sketches drafts, style gets revised back and forth, the final version gets cropped for each platform's dimensions — a long cycle with high communication overhead, and style alignment relying entirely on manual back-and-forth. Revising the mood means redrawing from scratch, and trying three directions means three times the work. This early concept-exploration stage is exactly where AI image generation can take over — quickly producing a batch of directional images to set the tone, before moving into refinement and final delivery.

Which model or tool handles which part of a game's promotional visuals? A quick reference table

The KV and the Steam header aren't produced end-to-end by a single model — the process splits into three stages, "base art → fix flaws → add text," each using the tool best suited for it:

The key point of this table is: don't force a single model to carry the whole process. Midjourney's strength is mood and stylistic range — asking it to render text is forcing it into something it's not good at. Flip it around and hand the tone-setting stage to an instruction-following model instead, and you lose that artistic feel. Once split into stages, each stage uses the right tool for the job, and the floor for quality across the whole set immediately gets much more solid.

Beyond the three stages, there's one more hidden workhorse: the reference image. Midjourney has a hard time precisely replicating a specific game's art style from text prompts alone, especially for one-of-a-kind styles like pixel art, a specific color palette, or a specific brushwork. Feeding it in-game screenshots, character art, or scene concept art as reference images to anchor the direction works far better than stacking a hundred adjectives — this step is the make-or-break move for avoiding style mismatches, and we'll dig into it in the hands-on section below.

What kind of game publisher or developer are you? Find your matching setup

Teams of different sizes and with different art resources apply this workflow differently — find the one that matches your situation:

The common baseline across all four types comes down to one rule: manage the base art and the text separately. Let Midjourney set the mood and feel, hand the game name and tagline to a model that renders text reliably, and never expect a single model to deliver the finished text version in one shot — especially in scenarios that need to pass Steam review or ship in multiple languages.

From zero to launch: what does the full workflow for a game's promotional KV look like?

1. Gather materials and set the direction (about 15 minutes): Pull together three to five in-game screenshots, character art pieces, or key scene concept art that best represent the art style, to use as reference images. At the same time, write a sentence or two capturing the KV's core message — world keywords, protagonist, mood (for example, "post-apocalyptic pixel wasteland, solo exploration, cold color palette"). This step sets the "art direction" — the anchor that keeps the whole set of images from drifting off-course.
2. Generate the KV concept base art (about 15 minutes): In Midjourney V7, at a 16:9 landscape ratio, upload the gathered game screenshots as reference images and pair them with mood prompts, generating 4 images per batch. The reference images pull Midjourney's style toward the game's actual art direction — check the batch of 4 for a composition and tone that both land. If none fit, adjust the reference image weight or swap the reference image and regenerate — don't rush ahead.
3. Finalize and refine (about 10 minutes): Pick the image with the best-matching direction from the batch of 4, and upscale it into a high-resolution base image. Check the image for local flaws — common issues in pixel art include rough edges, broken local structure, or color palette bleed. Box off any flawed areas and hand them to Nano Banana 2 for local inpainting to fix separately, outputting the final version at up to 4K, leaving the main subject and overall style untouched while only patching local spots.
4. Derive the Steam header (about 10 minutes): Once the KV's tone is set, move on to the store header. The header's main ratio is close to 460x215 landscape — use 16:9 or an even wider format description to match this squat frame, recomposing the KV's main visual into a horizontal header base, and leaving a clean spot reserved for the game name and title text. Generate 4 images per batch and pick the composition best suited for adding text.
5. Add the logo and title text (about 10 minutes): With the base art finalized, add the text last. Hand text elements — the game logo, game name, and platform tagline — to GPT Image 2, which renders text reliably and follows instructions well, placing Chinese or multi-language titles accurately into the reserved spot. Check each finished text version image by image for typos or text overlapping the main subject, then export at each platform's required dimensions for launch once confirmed clean.

A pixel-art title's KV came out photorealistic — how do you fix it? A real failure and recovery case

Last month I took on publishing for a pixel-art indie puzzle title and needed a promotional KV for the official site and social media. The game's art was that low-resolution, restrained-palette retro pixel style, very recognizable. Trying to save time, I skipped the reference image on the first pass and ran it straight off text prompts: Midjourney V7, 16:9 landscape, with "pixel art, retro, post-apocalyptic, exploration" written into the prompt, generating 4 images. The result was a complete miss — all 4 images came out as polished photorealistic digital paintings, with one landing squarely in that rounded Western-cartoon 3D look. The lighting was gorgeous, but it had nothing to do with the game's actual rough-grain pixel feel. If players walked into the store off that image, they'd feel misled the moment they saw the actual gameplay.

The problem was clear: Midjourney's stylistic pull is strong, and if you only give it text prompts, it drifts toward whatever it "thinks looks good" — the words "pixel art" alone weren't enough to override its aesthetic tendencies. The fix came down to a single move: anchor the art direction with a reference image.

Step one: I pulled real in-game screenshots and picked the three that best represented the art style — one of the protagonist in a pixel scene, one close-up of a signature item, and one distant-view level shot. I uploaded these three as reference images so Midjourney would generate off the game's actual art style instead of guessing at what pixel art should look like.

Step two: I reran the generation with the reference images attached, 16:9 landscape, 4 images per batch. This version nailed it immediately — the color palette, the grain, and the pixel block size all tracked the actual game, and the composition captured that solo-exploration feeling I wanted. The reference image turned the abstract phrase "pixel art" into this specific game's actual pixel style — that's the gap between it and pure text prompting.

Step three, the finishing touch. The chosen base image had one small local flaw: a patch in the bottom-left corner where the scene edge pixels blurred into a smudge with unclear structure. I boxed that area and handed it to Nano Banana 2 for local inpainting to fix it separately, leaving the style and subject untouched while only patching the blurred spot, outputting the final version at up to 4K. For the Steam header, I recomposed this KV into a horizontal base close to the 460x215 ratio, leaving space for the title, and handed the game name plus a "coming soon" tagline to GPT Image 2 to overlay — it rendered even the pixel-style logo lettering accurately based on the reference, far more reliable than asking Midjourney to spell it out. This mishap nailed down a rule I now follow strictly: for strongly identifiable art like pixel art or a specific color palette, use a reference image from the very first pass — don't gamble on Midjourney guessing right.

Pre-launch checklist: game promotional visuals

• Art style matches the actual game: The KV and header's art style must match the actual game visuals — a mismatch is a direct source of bad reviews.
• Use a reference image first for strongly identifiable art: For pixel art, hand-drawn styles, specific color palettes and the like, upload in-game screenshots as reference images to anchor the direction from the very first pass.
• Handle text elements separately: Give the game name, logo, and tagline to a model that renders text reliably — don't force the mood base-art model to spell text out.
• Match the aspect ratio to the platform: KV art runs 16:9 landscape; match the Steam header to its roughly 460x215 landscape frame using a wide-format description.
• Leave a spot for text: When generating the base art, reserve a clean, unobtrusive area for the title — don't wait until the text stage to discover the subject is blocking everything.
• Box off and fix local flaws: For rough edges, broken structure, or color palette bleed, use local inpainting to patch just that spot without touching the overall style.
• Check every finished text version individually: Especially for multi-language launches, check character by character — catch typos or text blocked by the subject before exporting.

When does an aggregator platform not make sense?

For established teams with ample art resources, a mature outsourced art pipeline, and ready-made promo-grade KVs, AI in the early concept stage is more of a supplement than a primary tool — this workflow won't add much for them. And if you already have a Midjourney direct subscription with enough usage quota, just keep using that directly — paying twice for the same thing doesn't make sense. Direct access to the original provider requires an overseas network environment and an overseas account setup, which this article won't go into. Worth clarifying: a so-called "domestic gateway to overseas models" is, at its core, an aggregator platform connecting original-provider models like Midjourney V7 for use within mainland China — the model capability belongs to the original provider, while the platform provides stable access, a unified account, and credit-based billing. The method itself — anchoring art direction with reference images, keeping base art and text separate — has nothing to do with which platform you use; wherever you generate images, style accuracy and clear division of labor are both worth sticking to.

• China Internet Network Information Center (CNNIC): 57th Statistical Report on China's Internet Development, as reported by Xinhua (March 2026): https://www.news.cn/tech/20260302/66c4ab06b6f34f8d806b416b3acc9f0b/c.html, official site: https://www.cnnic.net.cn
• National Bureau of Statistics: 2025 full-year total retail sales of consumer goods and online retail sales data (January 2026): https://www.stats.gov.cn/sj/zxfbhjd/202601/t20260119_1962345.html
• Flux Art official site: https://flux-art.ai and https://flux-art.cn

Flux Art is an all-in-one AI visual generation workbench: one account aggregates 50+ leading global image and video models (GPT Image 2, the full Nano Banana lineup, Midjourney V7, Grok Imagine, Grok Video 3, Seedance 2.0, and more), with direct, stable access within mainland China, output up to 4K with no watermark and commercial use allowed, plus 20K+ prompt templates and 150+ vertical-specific Agents. The operating entity is MORNING STAR INDUSTRY LIMITED. Official site: https://flux-art.ai and https://flux-art.cn. Note: Flux Art is an aggregator platform, not FLUX.1 or any single model from Black Forest Labs; each model's capability belongs to its original provider, made accessible within mainland China through Flux Art. Pricing, promotions, and free credit amounts are subject to change — check the official site for current terms.