---
title: UniVideo-train
canonical_url: "https://www.modelscope.cn/models/yansongwei/UniVideo-train"
md_url: "https://www.modelscope.cn/models/yansongwei/UniVideo-train.md"
repository: yansongwei/UniVideo-train
last_updated: 2026-08-17
pipeline_tag: text-generation
tasks:
  - text-generation
library_name:
  - pytorch
frameworks:
  - pytorch
downloads: 0
stars: 0
---

# UniVideo-train

> UniVideo-train - yansongwei 在 ModelScope 开源的模型。LoRA checkpoints fine-tuned for CAP accident-oriented text-to-video generation. Only the adapter weights required for inference are published; optimizer/training states, datasets, logs, smoke-test outputs, and…

- **Repository**: yansongwei/UniVideo-train
- **Tasks**: text-generation
- **Downloads**: 0
- **Stars**: 0
- **Last updated**: 2026-08-17

Source: https://www.modelscope.cn/models/yansongwei/UniVideo-train

---

# UniVideo CAP LoRA

LoRA checkpoints fine-tuned for CAP accident-oriented text-to-video
generation. Only the adapter weights required for inference are published;
optimizer/training states, datasets, logs, smoke-test outputs, and base-model
weights are excluded.

## Base components

- UniVideo with Qwen2.5-VL-7B conditioning
- HunyuanVideo transformer backbone

The matching UniVideo, Qwen2.5-VL-7B, and HunyuanVideo base checkpoints are
required for inference.

## Training configuration

- Training steps: 3,246
- Checkpoint interval: 500 steps, plus the final checkpoint
- Per-rank batch size: 1
- Learning rate: 1e-4
- Optimizer: AdamW
- LoRA rank: 8
- LoRA alpha: 16
- LoRA dropout: 0.05
- LoRA target modules: `to_q`, `to_k`, `to_v`, `add_q_proj`, `add_k_proj`,
  `add_v_proj`, `to_out.0`
- Trainable LoRA parameters: 13,172,736
- Training dtype: BF16; saved adapter dtype: FP32
- Training resolution: 288 x 160
- Frames per clip: 33
- Seed: 42

## Checkpoints

The repository includes `lora.pt` for steps 500, 1000, 1500, 2000, 2500,
3000, and 3246. Each file contains 536 LoRA tensors. The final adapter is:

`checkpoint-0003246/lora.pt`