HeyHomees Lab Notes · No. 8 · Tencent · Published October 23, 2026
By HeyHomees Newsroom ·
In 30 seconds
- FAPL makes MLLM latents usable by a Video DiT for cinematic video generation.
- It adds three training stages without changing the model architecture.
- Reported gains include Motion Order 61%, ∩-Acc 58%, and FaceSim-Arc 0.523.
What is it?
FAPL (Factor-Aware Predictive Learning) is a training framework that enables multimodal large language model (MLLM) representations to be used by a Video Diffusion Transformer (DiT) for cinematic video generation. It targets temporal inference from limited inputs and factor-aware control of character, scene, action, and camera.
How does it work?
FAPL uses three sequential stages: (1) freeze the MLLM and connector while training the DiT with only MLLM latents as content condition; (2) keep the MLLM frozen and train the DiT to perform factor-span replacement on paired videos for factor-aware editing; (3) freeze the trained DiT and optimize the MLLM and connector through it using text-only and text-plus-complete-video observations sharing one video target. No architectural changes are made.
How well does it do?
On the UniVideo (13B DiT) backbone, FAPL improves Motion Order from 48% to 61%, ∩-Acc from 42% to 58%, and FaceSim-Arc from 0.451 to 0.523, according to the authors' experiments.
By the numbers
- Motion Order: 61%
- Intersection Accuracy (∩-Acc): 58%
- FaceSim-Arc: 0.523
The honest read
The reported numbers are self-reported improvements by the authors on their own experimental setup; the sources do not provide external benchmarks or comparisons. The figures show relative gains over the UniVideo baseline but do not indicate absolute performance on downstream tasks or real-world video quality. No independent evaluation is presented.
What we don't know yet
- Exact hardware requirements for training or inference are not specified.
- The size of the model checkpoints or dataset used for training is not stated.
Sources
- Paper (arXiv abstract) (primary)
- Code, GitHub
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.