An AI concept art pipeline for game studios, model by model

Concept art has two modes and they want opposite things. Exploration wants volume, surprise and cheap mistakes. Lock-down wants the same face, the same silhouette and the same palette across forty images that three other departments will build from. A pipeline that only does one of those is not a pipeline. Here is how we split the work across the three image models, Characters and Elements, with the arithmetic at the end. The summary version is on the concept art use case.

Three models, three jobs

All three image models cost 10 credits per image, all three take a negative prompt, CFG control and reference photos, and all three batch 1 to 4 images at a time across 1:1, 3:4, 4:3, 9:16 and 16:9. What differs is what they are good at:

• Image Art — painterly and prose-faithful. It follows long descriptive prompts, which makes it the exploration model. If your art lead writes in paragraphs, this is where those paragraphs go.
• Image Anime — anime, manga and illustration. The model for stylized projects, cel-shaded characters and splash art.
• Image One — photoreal flagship at 1024×1024. Less for the concept itself than for the boards around it — material studies, environment reference, the photo-real plate an artist paints over.

Mixing them on purpose is the point. A painterly key art pass in Image Art, a photoreal material board in Image One, and the two read as one project as long as the palette wording stays the same.

The exploration pass

Exploration is a volume exercise, so run it as one: batch four, wide 16:9, one paragraph per idea, and do not fix anything yet. CFG scale is the dial that decides how literally the model follows your words — nudge it up once the brief is locked and leave it lower while you still want the model to argue with you.

The negative prompt is where the studio's house rules go. "text, watermark, logo, signature" belongs in every environment pass unless you enjoy cloning out invented runes.

Here is a complete Image Art prompt from a look-dev sprint for a rain-soaked harbor level:

A drowned harbor town at low tide seen from a broken pier, hulls of fishing boats tilted in grey mud, rope bridges strung between roofs above the waterline, lanterns burning low in the fog, heavy rain falling through a cold green-blue palette with one warm lantern accent, painterly brushwork, matte painting composition with a low horizon, wide 16 by 9 frame.

Long, descriptive, no bullet lists — that is the register Image Art rewards. If a prompt comes out thin, Enhance rewrites it into a director-grade version for free.

Characters keep the hero the same person

A hero design that drifts between images is not a design, it is a mood board. Use Characters — the persistent cast, each with a name, a persona, reference photos and a voice. The reference photos hold the same face across images and clips, and the character is @mentionable in any prompt, so a turnaround is a dozen prompts that all start with the same @name.

That mechanism is a reference image with a memory, and it is the difference between "draw this knight again" and "draw this knight from behind". Pin the hero early, then iterate on armor, weapon and silhouette around a fixed face.

Elements keep the world the same world

Elements are the same idea for everything that is not a person — reusable places, objects and styles, up to 50 per account, @mentionable and usable in both image and video prompts. One Element per location, one per faction's material language, one per art-direction style, and the environment team stops re-describing the harbor every morning.

This is also the piece that scales to a team. An Element written once by the art lead is a house style that everyone's prompts inherit, rather than a PDF nobody opens.

What a sprint costs

Images are 10 credits each, so the arithmetic is unusually honest:

• An exploration morning — 10 ideas, batched 4 at a time, is 40 images = 40 × 10 = 400 credits
• A hero turnaround — 12 angles and outfit variants = 12 × 10 = 120 credits
• A full look-dev sprint — 200 images = 200 × 10 = 2,000 credits

At the entry pack rate of $10 for 1,000 credits, that sprint is about $20. On plans, Creator ($49 for 6,000 credits) is 6,000 ÷ 10 = 600 images a month; Studio ($149 for 20,000) is 2,000 images; Team ($249 for 36,000 credits, 25 seats) is 3,600 images across the studio, or 36,000 ÷ 2,000 = 18 look-dev sprints. Annual billing takes 20 percent off, credits roll over forever and failed renders refund themselves, so an experimental month is not a wasted one.

One more line item worth knowing: a key frame you like can be animated. Animate turns a still into an image-to-video source, and a 5-second clip on Motion One is 100 credits — which is a moving mood piece for a pitch deck at the price of ten images.

Feeding the rest of the studio

Every render is saved automatically in your library with its full recipe, so the variant someone asks for in three weeks is a Remix rather than a re-derivation. Members of a workspace share one wallet, so the art team spends from one budget instead of five personal cards.

When the volume outgrows clicking, the same models are available through the API — images/generations takes the same prompt, model, negative prompt and aspect ratio you have been using in the studio, and the image docs list the parameters. A nightly job that renders forty variants of the week's brief is about fifty lines of script.

FAQ

Which model should a stylized project default to?

Image Anime for anime, manga and illustration work, and Image Art when the brief is painterly or written as prose. Both are 10 credits per image, so run the same prompt through both once and let the art lead pick — the comparison costs 20 credits.

How do I keep a character consistent across an entire deck?

Create the character once with reference photos, then @mention it in every prompt. The same references carry the face across images and clips. Keep the palette wording identical between prompts and the deck reads as one hand.

Is mature content allowed in concept work?

Mature imagery of adult, fictional characters is allowed within the acceptable use policy — adults only, fiction only. Sexual content involving minors and real identifiable people without consent are banned absolutely, and blocked requests return an error and are never charged.

Got a brief and an afternoon? Open the image studio, batch four, and keep the two that surprise you.