Testing Video Coherence: Torturing Seedance 2.5 with a Kitchen Timer
Brave_Pressure_9886 · reddit · 2026-08-03
ByteDance's Seedance 2.5 supports up to 30 seconds of video generation. To test the model's object state continuity and physical logic over longer clips, the author designed a rigorous testing protocol.
- The Scene: A digital kitchen timer is set to 20 seconds with a locked camera. A person folds a towel, sets it aside, and stops the timer when it beeps at zero.
- Evaluation Criteria: Digits must count down accurately, audio must sync with actions, and the folded towel must not randomly change state. Rendering errors (e.g., a distorted '8') are scored separately from logic failures (e.g., countdown stalling).
- Controlled Experiment: Using the model's reference control feature, the author will run the same prompt with minimal vs. heavy reference guidance multiple times to see if references actually improve continuity or if success is just luck.
- Model Comparison: Plans to run the identical scorecard on LingBot-Video to evaluate its physical rationality. If both pass easily, difficulty will increase (e.g., temporarily moving the timer out of frame) to test state memory limits.
More from Multimodal
- ByteDance's Dreamina Launches Seedance 2.5 with Multi-Shot Consistency — Div_pradeep · 2026-08-03
- Open Video Models: License Restrictions and VRAM Requirements Compared — Mysterious_Sign_9501 · 2026-08-03
- 30B Video MoE Quantizations Tested: Most Users Should Wait — EntireBig7258 · 2026-08-03
- GPT Image 2 Prompt for Surreal Editorial Portraits — SimplyAnnisa · 2026-08-03
- Midjourney Workflow: Crafting Cohesive Nighttime Visuals with SREF — tisch_eins · 2026-08-03
- Seedance 2.5 Demo: Generating a 32-Creator Brawl from a Single Prompt — azed_ai · 2026-08-03