Wan 2.2 LoRA training guide, scoring mechanics clarified

Team detailed the mining playbook: deploy stock Wan 2.2 I2V-14B first, then iterate on weak dimensions via LoRA fine-tuning (rank 32, ~$25-50 per run on rented H100). Scoring uses winner-take-all with 5% dominance threshold over 100-task windows; rank 1 requires sustained ≥5% pass-rate lead, not a one-off bump. First-frame substitution, scheduler swaps (UniPC/DPM++), and negative prompts are the real differentiators; the model is a floor, not a ceiling.

• •Deploy Wan 2.2-I2V-A14B 14B as baseline; smaller TI2V-5B will cap quality scores.
• •LoRA fine-tune (rank 32) on curated I2V pairs; overfitting kills prompt adherence (25% of rubric).
• •Single Pro 6000 96GB with FP8 quantization recommended; multi-GPU only if necessary.
• •5% pass-rate margin required to flip rank 1; must sustain it over rolling 100-task window.
• •H100 80GB rental ~$2-3/hr; budget $100-200 total for competitive LoRA via 2-3 iteration cycles.