🚀 適用於Wan2.1 14B T2V的廢棄場所LoRA
本項目基於Wan2.1 14B T2V模型訓練了一個LoRA,能夠生成廢棄場所相關的視頻,為文本到視頻的轉換提供了新的可能性。
🚀 快速開始
模型文件下載
推薦設置
- LoRA強度:1.0
- 嵌入式引導比例:6.0
- 流偏移:5.0
觸發詞
關鍵觸發短語為:abandoned places
ComfyUI工作流
此LoRA可與 Kijai的Wan視頻包裝器工作流 的修改版本配合使用。主要修改是添加了一個連接到基礎模型的Wan LoRA節點。
修改後的工作流見上方下載部分。
✨ 主要特性
- 基於Wan2.1 14B T2V基礎模型進行訓練。
- 在不同的對象和場景類型中能保持一致的生成效果。
- 提供簡單且易於調整的提示示例。
📦 安裝指南
暫未提供明確的安裝步驟,可參考上述下載和設置信息進行操作。
💻 使用示例
基礎用法
以下是一些示例提示及其對應的輸出視頻:
- 提示文本:abandoned places A steady zoom-out tall abandoned building covered in vines and trees stands in the middle of an abandoned city. The sky is overcast and the air is thick with fog. The city is mostly obscured by the fog, but you can see some other buildings in the distance. There is a sense of decay and abandonment in the image.
輸出視頻:example_videos/1.mp4
- 提示文本:abandoned places A steady zoom-in to An old, rusty car, partially obscured by vegetation, sits in the middle of a dense forest. The car's body is a faded light blue color, with a rusted hood and roof. The car's door is open and the interior is dark and dusty. The windows are all broken and the seats are torn. The car is surrounded by tall trees and thick bushes. The ground is covered in leaves and debris. The overall atmosphere is one of decay and neglect.
輸出視頻:example_videos/2.mp4
- 提示文本:abandoned places A grand, abandoned mansion, with peeling paint and broken windows, stands in a grove of overgrown trees. The front steps are crumbling and the driveway is cracked. The sky is overcast.
輸出視頻:example_videos/3.mp4
- 提示文本:abandoned places A steady zoom-out from the center of an abandoned industrial complex, where a single rusted smokestack stands tall against the cloudy sky. The camera moves back, revealing a network of crumbling buildings and broken glass windows. The sound of distant thunder rumbles as the wind kicks up dust and loose debris.
輸出視頻:example_videos/4.mp4
高級用法
對於提示,可參考上述示例提示,這種提示方式似乎效果很好。
📚 詳細文檔
模型信息
模型權重以Safetensors格式提供,詳見上方下載部分。
訓練詳情
屬性 |
詳情 |
基礎模型 |
Wan2.1 14B T2V |
訓練數據 |
基於6分鐘的視頻進行訓練,該視頻由98個廢棄場所的短視頻片段組成(每個片段單獨標註) |
訓練輪數 |
50 |
額外信息
訓練使用了 Diffusion Pipe for Training。
致謝
特別感謝Kijai提供的ComfyUI Wan視頻包裝器,以及tdrussell提供的訓練腳本!
📄 許可證
本項目採用Apache-2.0許可證。