Generating and Evaluating Custom Pixel Characters with StripeFoundry
A custom pixel character must keep its distinguishing marks across poses on a grid small enough for an artist to edit cell by cell. We present StripeFoundry, a pipeline that specifies the lattice before generation, projects an editor's drawing onto native cells, and tests each pose before using it as an animation anchor. We compare it with ordinary Stable Diffusion image-to-image sampling on the same 25 pose references, including an identical projection for both generators. At strength 0.65, median boundary alignment was 0.478 for diffusion and 0.890 for the guided editor, while median reference silhouette IoU favored projected diffusion (0.912 versus 0.770 for the final sprites). Across five canine character sheets, 19 of 25 poses cleared the pipeline's checks on the first attempt and 23 cleared them within three. Of 34 candidates, six failed geometry, three failed eye visibility and two failed symmetry. All 25 selected sprites had binary alpha, but two failed poses were retained and appeared directly in four of 30 animation calls. These measurements distinguish an editable asset from a verified pose and locate the limits of both guarantees. Hyeon Soek Cho, Seo Young Kim, and Pdo Hyun Park contributed equally as co-first authors. Minseop Choi is the corresponding author.
Authors
- Minseop Choi
- Seo Young Kim
- Hyeon Soek Cho
- Pdo Hyun Park
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-20
- DOI
- https://doi.org/10.5281/zenodo.24378573
- Primary Topic
- Generative Adversarial Networks and Image Synthesis
- Type
- preprint