Workflow: drafting tileset concept art with an AI image generator

[marlowfinch42]

Posting some notes on the draft stage, before any real pixel work starts. The loop that has worked:

1. Fix the lighting and view angle in the brief first: tile footprint, camera angle, light direction, palette. Isometric tiles fall apart quickly if the light direction drifts between tiles.
2. Feed a rough sketch or an existing tile as an image-to-image / sketch reference instead of relying on text-to-image alone. Keeping the silhouette readable matters far more than detail.
3. Batch variants per terrain type, then pick one and hand-trace it. Almost nothing generated survives at full detail; the useful part is the colour ramp and the shading direction.
4. Check every candidate at the real tile scale. What reads well at large sizes is usually mud once downscaled.

Generator used for the drafts: Seedream 5.0 Pro -- browser-based, with text-to-image, image-to-image, sketch-guided, multi-reference and batch runs. Not affiliated, and the workflow is the point: swapping the generator does not change steps 1-4.

One caveat worth stating up front: generated drafts are reference material, not final assets. Freeciv art has to be redistributable under the project's licence terms, so treat generator output as something to re-draw rather than something to ship.

Curious how others handle the sketch-to-tile step, especially for coastal and river transitions where the adjacency rules bite.