Answers for developers, publishers and site owners on monetizing and distributing web games.
336 posts
Use Draco (KHR_draco_mesh_compression) for maximum geometry size reduction, or Meshopt (EXT_meshopt_compression, via gltfpack) for faster decoding with less CPU cost. Both need a matching decoder in your loader; three.js and Babylon.js support each with version-specific setup.
Fix the palette before you generate: build a reference sheet or color-locked style/seed image, feed it to every prompt or img2img pass, then post-process outputs (palette swap or indexed color reduction) so all assets share the same swatch set before they hit your sprite sheets.
Store textures as KTX2 with Basis Universal supercompression, ship them small over the network, then transcode in the browser to a block-compressed format the GPU supports (ASTC, ETC, BC) before uploading - this shrinks both download size and GPU memory, since WebGL keeps compressed textures compressed in VRAM instead of expanding them.
Use PNG for sprites that need transparency, packed into sprite sheets to cut HTTP requests. Use JPG or WebP for opaque background art where file size matters more than alpha. GPU compressed formats (KTX2/Basis) help VRAM, not download size.
A sprite sheet JSON needs one PNG plus a data file listing, per frame, its x/y/width/height in the sheet, and usually a name or tag. Export this directly from Aseprite or an AI sprite tool rather than hand-writing it, and load it with your engine's own frame/atlas loader.
Name the asset type (sprite, not scene), the view angle, the pixel grid or art style, pose or action, and transparent background explicitly. Tools that offer a fixed image-type picker (sprite, tiling sprite, texture) get more consistent results than free-text prompts alone.
Use a LoRA when you need a fixed art style trained on your own reference set, IP-Adapter when you need the same face or character reused across poses from one or two images, and ControlNet when you need to force a specific pose, silhouette or layout. Most consistent asset pipelines combine two of the three.
A sprite sheet is a grid of equal-size frames, usually one animated object, sliced by fixed width/height. A texture atlas packs many differently sized images into one sheet plus a JSON file mapping names to positions - used for whole tilesets, UI, or many objects in one texture.
Send less to the model: truncate or summarize old conversation turns, cache stable context (system prompt, NPC lore), use cheaper models for routine lines, and batch non-urgent requests. Most cost in chat-based NPCs comes from re-sending full history on every call, not from the reply itself.
Never let an LLM's raw text run as code. Have the model return structured output (a function call or JSON matching a schema), validate it against a whitelist of allowed actions and parameters, then let your game code - not the model - execute the action.