Open models: 4 research artifacts on HuggingFace

We publish research artifacts from our production and R&D work as open-weight models on HuggingFace. These are not demos — each model was either trained for a shipping product or validated a specific research hypothesis.

cosmic-deco-flux-lora — 1970s NASA retrofuture aesthetic (Don Davis, Rick Guidice). Trained on 263 public-domain NASA images. Used in production at gpu.social. 88 MB, rank 32.

ukiyo-e-flux-lora — Edo-period Japanese woodblock print style (Hokusai, Hiroshige). Flat color planes, contoured linework, traditional pigment palette. 176 MB, rank 64.

jppy-logo-flux-lora — logo and brand-mark generation. Trained on 15,815 logos. Live in production at gpu.social for automated logo design. 2.3 MB, low rank.

yuka-dora-v1 — DoRA adapter on Qwen3-8B for personal voice replication. First server-trained DoRA that achieved 100% blind A/B win rate vs stock Qwen3-8B, with zero catastrophic forgetting. 63 MB, rank 16. Gated, CC BY-NC 4.0.

All models include full training configs, evaluation methodology, and reproduction instructions in their model cards. Need a custom model for your use case — reach out at [email protected].