ReentryMem
inference checkpoint for dynamic object reentry in video models
2 min readupdated
Built on Wan-AI/Wan2.2-TI2V-5B, this trained checkpoint tackles dynamic entity reentry in video. It merges DiT (Diffusion Transformer) weights and camera conditioning parameters at step 1496 in BF16 format. Aimed at research-grade video generation with precise camera and object positioning control, it requires loading the text encoder, tokenizer, and VAE separately from the base model. The release is geared toward researchers and developers working on advanced video generation.
Key features
- Merges BF16 DiT weights and camera conditioning parameters at step 1496.
- Supports the single-entity variant for the virtual scene reentry task introduced in the paper.
- Uses calibrated sampling across entity regions via the localized mode.
- Default generation pipeline handles 77 observed and 76 future RGB frames at 832x480 resolution.
What you can do
- Generate videos where a specified object dynamically reenters the frame based on a text description.
- Run video generation with camera pose control using the project repository.
Example
Use the official Larry-077/ReentryMem repository to run custom inference with the single-entity checkpoint and a text description of your target entity.
Read more
DetailsBase model: Wan-AI/Wan2.2-TI2V-5B
Format: BF16 DiT checkpoint (step 1496)
Default generation settings: 77 observed frames, 76 future frames, 832x480, 15fps, 50 steps, CFG5
Requires loading the text encoder, tokenizer, and VAE separately from the base model










Comments