DATASET

• 2231 Gemeindebau images, self-photographed across Wien, 2025-2026.
• Flux.1-dev, Ostris AI-Toolkit, single RTX 4080 16GB.

WHAT THE LOSS CURVE HIDES

v1 flatlined at 0.70 for ten thousand steps. The model learned nothing. Three parameters separated it from v3: a trigger word, rank 64 instead of 32, and switching the training mode from balanced to style. That last one did most of the work — it stops the model splitting its effort between content and style and lets it concentrate on aesthetic transfer alone.

v3 reached 0.46, the lowest number in the series. It is also the most misleading. v3 trained across all noise levels including the easy ones, which pulls the average down. v5 and v6 concentrated on the low-noise timesteps where facade materiality actually lives — higher loss, better images. The metric was measuring the wrong thing.

RUNS

• v4 collapsed. The optimizer decayed its own learning rate to nothing within the first steps and the run never recovered.
• v5 and v6 produced the two usable models — one for subtle interventions, one for deeper transformation.
• Both learned colour more strongly than structure. Caption dropout was 0.05. The next run raises it to 0.30.

v3 CHECKPOINTS

• 2500 — EMA global minimum. Cleanest convergence, best style/prompt balance.
• 5500 — second-best EMA. More style absorption, stronger imprint.
• 4500 — solid mid-run, tested against 5500 for texture over-commit.
• 6000 — skipped. Raw-loss spike to 1.01, instability bump.

WHAT DID NOT HAPPEN

• LTX-Video LoRA never started. Time ran out before the video stage.