EgoX mix LoRA — rank-128 QLoRA (single 24 GB GPU reproduction)

A LoRA adapter for Wan2.1-I2V-14B-480P reproducing EgoX (exocentric → egocentric video generation) on a single 24 GB GPU (vs the paper's 4× H200). Trained with the paper's Geometry-Guided Self-Attention (GGA), scaled down to fit the hardware budget.

exo input | our vitS ego-prior | ego pred (our prior) | ego pred (shipped prior)

Training

• Adapter: LoRA rank 128 / α 128 on the Wan-14B transformer (NF4 4-bit base + 8-bit AdamW)
• Data: 552 clips, all 17 EgoX domains (Ego-Exo4D); 495 train / 57 val
• Resolution: 49×176×704 · LR 2e-5, constant-with-warmup · effective batch 4
• 16 epochs / 1,984 steps, held-out val loss 0.247 → 0.173 (checkpoint-1984)

Use

Load as an unfused PEFT adapter onto an NF4-quantized Wan2.1-I2V-14B with GGA enabled (see the reproduction repo: EgoX/infer_nf4.py --use_GGA).

Caveats

Scaled-down repro: 176-res output, NF4-r128, 552 clips (vs the paper's 3,510). Image-criteria scores trail the paper (11 vs 16 PSNR) due to resolution + budget, not pipeline fidelity.