GUIDE

What is the difference between an AI image generator and an AI game asset workspace?

An AI image generator returns a picture. A game asset workspace returns a typed asset inside a game project: it keeps the character it belongs to, the animation it is a frame of, its palette, its tags and its scene coordinates, and it exports Tiled JSON, sprite sheets and tileset ZIPs. Coding agents can read that structure over read-only MCP; they cannot read a folder of downloads.

PX-Art is not another AI image generator. It is a workspace that connects AI generation, art-asset management and game engineering, and it produces assets that both humans and coding agents can understand. Here is what that means concretely.

Four differences that matter in a real project

Where the difference shows up first

Consistency is enforced, not requested

An image generator re-decides everything each call. A workspace pins the things that must not change: a palette extracted from the character's base sprite and reused by every action, a body-size tier normalized against a proportion reference image with horizon and head-top lines, and a cached turnaround still that anchors every frame of a non-front facing. Five facings are generated and three are mirrored, so left and right cannot disagree. Details are in keeping AI-generated characters consistent.

Animation is a sequence, not a batch of pictures

Input is a base sprite, an action, a facing, a frame count and a loop flag. The pipeline composites onto a key-color canvas, generates a short clip, decodes every frame, samples by delivery semantics, keys out the background, crops the whole sequence against one union bounding box so movement is preserved, assembles a strip with one scale and baseline, applies palette post-processing, and runs quality checks for facing drift, inter-frame coherence and ground contact. Looping actions are sampled without a duplicated end pose. You cannot assemble that from independent image calls. See the pixel art animation generator.

Tiles are cut, not guessed

Square tilesets are drawn inside a magenta hollow guide grid and cut along the same fixed coordinates, so slicing is deterministic rather than inferred from content. The export is a ZIP with atlas.png, tileset.json, tileset.tsx and tileset.tmj, each tile carrying its role and source cell. Walkthrough: generating a tileset for Tiled, and the tileset generator.

Scenes keep coordinates

An arranged scene exports as Tiled JSON with orientation set to orthogonal or isometric, tile layers as row-major gid arrays where 0 is empty, an image-collection tileset with firstgid of 1, and layers carrying both x/y grid coordinates and offsetx/offsety canvas pixel origin. Non-tile nodes export as an objectgroup where every object has px_art_node_id, px_art_z, px_art_asset_id, px_art_asset_file and px_art_tags. That is the difference between "here is a picture of a level" and "here is a level".

Agents can query it

The MCP server is named perfectpixel-scene, speaks JSON-RPC, and implements protocol version 2025-11-25 while remaining compatible with 2025-06-18. It is read-only — there are no write, generate or delete tools — and exposes twelve tools:

list_games            list_game_structure   list_characters
list_asset_families   search_assets         preview_assets
get_asset_downloads   get_scene_bundle      get_scene_metadata
get_nodes             get_scene_region      analyze_layout

resources
  scene://{sceneId}/bundle        scene://{sceneId}/composition
  scene://{sceneId}/metadata      scene://{sceneId}/tiled
  asset://{id}                    asset://{id}/image

Tokens are issued per user and scoped to that user's own games; every call verifies subscription status and is written to an audit log. It works with MCP-capable coding agents such as Cursor, Claude Code and Codex. More at Agent / MCP, MCP game assets and AI game assets for coding agents.

How to tell which one you need

1. Count the images you need of the same subjectOne key art image: an image generator is fine. Twelve actions across eight facings for six characters: you need enforced consistency, or you will spend the saved time on retouching.
2. Ask whether position mattersIf the output has to land at a grid cell, on a layer, at a z-order, you need something that stores coordinates and exports them.
3. Ask who consumes it nextIf a coding agent or a build script reads the assets, they need queryable structure and stable ids, not a folder of PNGs.
4. Ask what happens in month threeA new character added later has to match the old ones. That is a project-level style key, a palette lock and a body-size tier — properties of a workspace, not of a prompt.

What an image generator still does better

Being specific cuts both ways. If you want one illustration, a mood board, a piece of concept art or a wallpaper, a general image generator is the shorter path — you do not need modules, scene coordinates or an MCP token to make one picture. The workspace pays off when assets have to agree with each other over time and be consumed by something other than a human eye.

Limits and caveats

• Structure does not make generation deterministic. Results are probabilistic and may not match your description; review assets before shipping them.
• MCP access is strictly read-only. Agents can list, search, preview and download; they cannot generate, edit or delete. Any subscription payment or top-up credit purchase permanently unlocks read-only MCP.
• The video-based action pipeline depends on the model obeying a key-color background and a fixed camera. When it does not, the run fails and asks for regeneration instead of delivering broken frames.
• The Arrange workspace needs a large canvas and precise mouse work; on a phone you can browse assets and take notes only.
• The final assembled prompt is internal and never returned to the user interface or the API.
• There is a thin Godot-side adapter. Unity is not supported and is not claimed.
• Legally generated assets can be used commercially, but you still need to check third-party rights, applicable law and AI provider restrictions. AI output is not guaranteed to be exclusive or registrable for copyright.
• Checkout is still under review and not open yet. New accounts start with 20 credits; a static image costs 5 credits, a frame-based action 10, a video-based action 18, an effect 15 and a UI kit 25.

FAQ

Related