A LoRA trained on all 100 plates of Kunstformen der Natur. Scroll to explore the training run and generated plates directly — drag, scrub, click.
Drag to scrub through the run. Loss drops for ~50 epochs, then goes flat — the remaining ~74 epochs bought almost nothing.
step 750 · converging
Captions never mentioned background color, so the model guessed. Fixed by measuring each plate's real background from its pixels — a clean, wide split, not a guess.
Two rounds of fine-tuning later (including one that doubled pale-example exposure for 40 more epochs), here's the honest result on the 12-plate eval sample — no net gain between rounds, some plates just traded places:
6/11 correct background color (1 plate excluded — full outdoor scene, not a flat-background plate)
Same caption the model trained on. A win, a nuanced case, and an honest miss — not cherry-picked.
Tafel 84 — Navicula, Diatomea. ✓ correct black bg — a genuine win.
Tafel 26 — Carmaris, Trachomedusae. ✓ correct pale bg, but the jellyfish rendered as an abstract cage.
Tafel 42 — Ostracion, Ostraciontes. ✗ teal, not pale — didn't move across two fine-tuning rounds.
Prompts that don't exist in the training set. No ground truth — the question is plausibility.
