- 🔥🔥 News:
2024/12/15: 🚀 We are excited to release the LoRA fine-tuning code for CogVideoX 1.5 by Diffusers, designed specifically for image-to-video (image2video) tasks! This update brings significant improvements and new features to elevate your training experience. The full training startup code can be found in thefinetunefolder. Here's what's new:
-
💥 Bucket-based Multi-Resolution Training: Unlock unparalleled model adaptability and performance across videos of all resolutions. This groundbreaking feature boosts the model’s ability to handle diverse video qualities with ease!
-
⚡ Fixed RoPE (Relative Position Encoding) Configuration Error: We’ve optimized the position encoding mechanism, solving the error in the original CogVideo code, resulting in smoother training and higher-quality outputs. No more misconfigurations—just pure efficiency!
-
🔧 Corrected OFS Embedding Issue: Previously, OFS embedding was incorrectly set to None in the original code. Now, it’s properly configured for stability and precision, improving the overall reliability and robustness of the model.
- Multi-resolution Support 🖼️
- Accurate Positional Encoding 📍
- Correct OFS Embedding Setup 🔑
- Optimized Multi-GPU Fine-Tuning 💻
With these changes, you can now start training with just one command—simple, fast, and effective!
🚀 我们激动地宣布发布了基于 Diffusers 的 LoRA 微调代码，用于 CogVideoX 1.5，并支持图像到视频（image2video）任务！ 本次更新带来了重大的改进和新特性，训练启动代码可以在 finetune 文件夹中找到。以下是主要的新功能：
-
💥 基于桶的多分辨率训练：释放前所未有的模型适应性和性能，适用于各种分辨率的视频。此项突破性功能增强了模型对不同视频质量的处理能力！
-
⚡ 修复了 RoPE（相对位置编码）配置错误：我们优化了位置编码机制，解决了原始 CogVideo 代码中的错误，从而提高了训练效率和输出质量。不再有配置错误——只剩下高效训练！
-
🔧 解决了 OFS 嵌入设置为 None 的问题：原始代码中，OFS 嵌入错误地设置为 None。现在它已经正确配置，提升了模型的稳定性和可靠性，确保了模型的精准性。
- 多分辨率支持 🖼️
- 准确的位置编码 📍
- 正确的 OFS 嵌入设置 🔑
- 优化的多卡微调 💻
通过这些改进，您现在只需一条命令就能开始训练——简单、快速、有效！