Wan 3.0 Tested: Major Physics Upgrades & 30-Second Clips
Fresh-Resolution182 · reddit · 2026-08-12
The author compared Wan 3.0 and 2.7 image-to-video (I2V) capabilities using identical first frames, prompts, and seeds at 720p, summarizing the practical upgrades and drawbacks.
Key Improvements:
- Physics & Motion Logic: 3.0 handles collisions and shattering (e.g., falling glass, billiard break) much more realistically. Motion logic (like a ballet spin with a following skirt) is also cleaner.
- Stability: Better subject and background consistency during movement, with fewer artifacts compared to 2.7.
- Long-form Video: 3.0 supports 30-second clips (up from 2.7's 15-second cap) and maintains identity/scene consistency without mid-clip swapping.
Known Bug:
- Uncontrolled Camera Cuts: Even without camera movement in the prompt, 3.0 randomly cuts to another shot mid-clip, sometimes with a ghostly dissolve. This happens on plain 5s clips and worsens over time, firing multiple times in 30s runs, which ruins continuous takes.
Despite the cutting issue, the author plans to switch to 3.0 for tasks where it clearly excels, while keeping 2.7 for quick iterations requiring clean, unbroken shots.
More from Multimodal
- AI Video Generation Bottleneck: Full Regens Kill Iteration and Momentum — SucceededMind · 2026-08-12
- LTX-2.5 Live in ComfyUI: Introduces Dynamic Compute Budget Rendering — Comfy-Org · 2026-08-12
- Midjourney Officially Rolls Out Version 8.2 Update — azed_ai · 2026-08-12
- Boost LTX 2.5 Video Generation Speeds by 5x with Conv VAE Swap — desktop4070 · 2026-08-12
- Dev Builds macOS Local Image App 'Diffusion Bus' Using Claude Code — deckarep · 2026-08-12
- Alibaba Quark Open-Sources LiveAvatar: 14B Model for Real-Time Infinite Audio-Driven Avatars — tom_doerr · 2026-08-12