Tilesets and animation

Single-frame art is largely a solved problem at this scale. These two are not. This page is honest about where the automation stops.

Tilesets

A tileset LoRA exists and is worth using as a starting point: it takes a shape (rectangular, hexagonal, octagonal) and a view (isometric, oblique, top-down) and produces tiles in a consistent style.

What it will not reliably give you is true seamlessness — edges that match perfectly when tiled, in every direction, across every tile in the set. Expect to fix seams by hand.

A workflow that respects that:

1. Generate a style reference tile you like.
2. Generate the rest of the set conditioned on it, so the palette and lighting agree.
3. Quantise the whole set to one shared palette — see Pixel-perfect output . This alone removes most of the “these tiles are from different games” feeling.
4. Fix edges in an editor. Budget real time for this; it is not a rounding error.

Seamless textures (as opposed to tilesets) fare better, because there is no grid to respect — only wrap continuity.

Animation

Two routes, neither fully automatic.

Spritesheet LoRAs. The installed 4-direction walk-cycle LoRA emits walk frames plus jump and prone at 32×32. This is the most reliable path because the frame relationships were learned, not inferred at generation time. Its ceiling is the specific animations it was trained on.

Video models. Video-generation models with pixel-art adapters can produce animation frames, and there are community LoRAs for walk cycles and attack animations. They are large, slow on a 90 W card, and frame-to-frame coherence is the exact thing they are weakest at — which is unfortunately the whole requirement for sprite animation.

The honest recommendation: generate key poses with the methods in Character consistency , then do the in-betweens and the timing by hand in Aseprite. That is faster and better than fighting a video model into producing eight coherent frames, and you keep control of the animation’s feel — which is the part players actually notice.

Why this section exists

It would be easy to write a handbook implying the whole pipeline is automated. It is not. Knowing precisely which two stages still need a human is worth more than an optimistic page that sends you into a two-day fight with a model that was never going to get there.