VideoFace-X-LLM — ViT-only fine-tune (step 3406 of 3406)

InternVideo2.5-Chat-8B fine-tuned on the ExplainFace corpus (Gemini descriptions of CelebV-HQ videos) for the ACCV 2026 paper "When, Where, and Why: Hierarchical Causal Explainability for Video Face Understanding with MLLMs". This is the "Ours (ViT-Only)" row of the results table.

Training

• Trainable: vision encoder (lr 1e-5) + connector mlp1 (lr 4e-5); language model frozen.
• 1 epoch = 3,406 steps over the 27,304-clip identity-independent training split (Aak579/CelebV-HQ-identity-split20), micro-batch 1, global batch 8, AdamW, weight decay 0.01, 3 % linear warm-up then cosine decay, gradient clip 1.0, 8–32 frames per clip, seed 42, FSDP full sharding (xtuner unify_internvl2_train_r16.py). Full args: train_run_config.json.
• This checkpoint is the end of the epoch.

Test results (6,923-clip identity-independent test split)

Loading

The folder is a full HF checkpoint (trust_remote_code=True; the modeling_*.py / conversation.py files are copied from the base snapshot). For the JSON tasks the model was decoded with the training-style prompt prefix and a { prefill, as in the paper's inference scripts.