You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Thank you for your interest and for raising this question!
For Qwen2-VL-7B on DREAM-1K, we used the prompt: "Describe the video in detail." The evaluation strictly followed the official Tarsier2 pipeline (https://github.com/bytedance/tarsier), using GPT-3.5-turbo-0125 for AutoDQ scoring. Each video was uniformly sampled into 64 frames, with a per-frame pixel limit of 100,352. Slight variations in frame count or resolution may lead to minor performance differences, but the impact is generally negligible. A comparable result is reported for Qwen2-VL-7B in the Tarsier2 Technical Report (https://arxiv.org/abs/2501.07888, Table 1).
Activity
Kerin637 commentedon Dec 14, 2025
Thank you for your interest and for raising this question!
For Qwen2-VL-7B on DREAM-1K, we used the prompt: "Describe the video in detail." The evaluation strictly followed the official Tarsier2 pipeline (https://github.com/bytedance/tarsier), using GPT-3.5-turbo-0125 for AutoDQ scoring. Each video was uniformly sampled into 64 frames, with a per-frame pixel limit of 100,352. Slight variations in frame count or resolution may lead to minor performance differences, but the impact is generally negligible. A comparable result is reported for Qwen2-VL-7B in the Tarsier2 Technical Report (https://arxiv.org/abs/2501.07888, Table 1).