To ensure training quality, we preprocessed the full 1,000,000 videos from the official dataset and randomly sample during training to maintain data diversity. We adopt the official parameters with batch_size=16 and rank=32 to prevent catastrophic forgetting caused by excessively large rank.

The VBVR dataset contains 100 reasoning task categories, with ~10,000 variants per task, totaling ~1M videos. Main task types include:

Object Trajectory: Objects moving to target positions
Physical Reasoning: Rolling balls, collisions, gravity
Causal Relationships: Conditional triggers, chain reactions
Spatial Relationships: Relative positions, path planning

Model Details

Item	Details
Base Model	ltx-2.3-22b-dev
Training Method	LoRA Fine-tuning
LoRA Rank	32
Effective Batch Size	16
Mixed Precision	BF16

TODO List

Dataset Release Plan

Dataset	Videos	Status
VBVR-96K	96,000	✅ Released
VBVR-240K	240,000	✅ Released
VBVR-reinforce	240K+150K	✅ Released

LoRA Capabilities

This LoRA adapter enhances the base LTX-2 model for production video generation workflows:

Enhanced Complex Prompt Understanding: Accurately interprets multi-object, multi-condition prompts with detailed spatial descriptions and temporal sequences, reducing prompt misinterpretation in production scenarios.
Improved Motion Dynamics: Generates smooth, physically plausible object movements with natural acceleration, deceleration, and trajectory curves, avoiding robotic or unnatural motion patterns.
Temporal Consistency: Maintains object appearance, lighting, and scene coherence throughout the video sequence, reducing flickering and frame-to-frame artifacts common in generated videos.
Precise Timing Control: Enables accurate control over action duration, pacing, and synchronization between multiple moving elements based on prompt semantics.
Multi-Object Interaction: Handles complex scenes with multiple objects interacting simultaneously, including collisions, following, avoiding, and coordinated movements.
Camera and Framing Stability: Maintains consistent camera perspective and framing throughout the sequence, avoiding unwanted camera shake or unexpected viewpoint changes.

Training Configuration

Stage 1: VBVR Foundation (96K)

Config	Value
Dataset	96K VBVR general videos
Learning Rate	1e-4
Scheduler	Cosine
Batch Size	1 × 16 (gradient accumulation)
Optimizer	AdamW
Max Grad Norm	1.0
Target Modules	`to_q`, `to_k`, `to_v`, `to_out.0`, `ff.net.0.proj`, `ff.net.2`

Stage 2: VBVR Extended (240K)

Config	Value
Dataset	240K general videos
Learning Rate	1e-4
Scheduler	Cosine
Batch Size	1 × 16 (gradient accumulation)
Optimizer	AdamW
Max Grad Norm	1.0
Target Modules	`to_q`, `to_k`, `to_v`, `to_out.0`, `ff.net.0.proj`, `ff.net.2`

Stage 3: General + Hard Reasoning (390K)

Config	Value
Dataset	240K general videos + 150K high-difficulty reasoning videos
Learning Rate	5e-5
Scheduler	Cosine
Batch Size	1 × 16 (gradient accumulation)
Optimizer	AdamW
Max Grad Norm	1.0
Target Modules	`to_q`, `to_k`, `to_v`, `to_out.0` (FFN frozen)

Video Demo

Training Progress Comparison

Original Model

240K

390K

Dataset

This model is trained on the VBVR (Video Benchmark for Video Reasoning) dataset from video-reason.com.

Contact

For questions or suggestions, please open an issue on Hugging Face or contact the author directly.

zhazhazha/Ltx2.3-VBVR-lora-I2V

LTX-2 VBVR LoRA - Video Reasoning

Training Data