Training LoRA with 20 photos, Ostris creates a personalized Wan2.1 video model

March 9, 2025, California, USA - Open source model is making Wensheng video technology more and more interesting. AI developer Ostris (@ostrisai) recently shared the results of his Wan2.1LoRA model trained on the X platform using his own photos. With only about 20 photos and a home RTX4090 graphics card, it achieved amazing video generation effects. This result not only demonstrates the potential of open source technology, but also triggers a heated discussion in the X community on video LoRA training tools.

Ostris' amazing experiment

Ostris first showed off the training results of Wan2.114B LoRA in a March 7 X post. He wrote: "Wan2.114B LoRA training runs successfully on 24GB! Average 1.7 seconds per step at 480p resolution on 4090." He also shared a preliminary video that demonstrated the feasibility of this model on home hardware. Then on March 9, he further released a "low-cost proof-of-concept short music video", which was paired with lyrics he created and music generated by @SunoMusic. The digital image in the video was trained on his own photos.

Ostris said he used only about 20 personal photos, completing the process with the video LoRA training tool he developed. He lamented in the post: "I didn't expect the effect to be so good! I had a lot of fun." This experiment not only verified the powerful performance of Wan2.1LoRA, but also showed the possibility of ordinary users using open source tools to achieve high-quality literary videos.

Training Tool Open Source Sharing Video LoRA training tool developed by Ostris has become a highlight of this release. User @sundyme posted on March 9: "Video LoRA training is here, let you become the protagonist of AI video!" and shared Ostris' tool link. This tool supports Wan2.1 model, allowing users to train personalized video models with a small number of photos, significantly lowering the technical threshold.

X Users show a strong interest in this tool. @sundyme said: "The LoRA training tool developed by @ostrisai supports Wan2.1 and has amazing results." Community feedback shows that the efficient operation of the tool on consumer graphics cards such as RTX4090 allows more creators to try AI video generation at home.

Community response and technical significance

The response on X shows that Ostris’s results have sparked widespread enthusiasm. One user commented: "It's incredible to train such a realistic digital person with 20 photos!" Another user praised the potential of the open source model: "Wensheng videos are becoming more and more interesting, and open source allows ordinary people to play with AI."

Industry insiders believe that Ostris' experiment highlights the breakthrough of open source models in the field of video generation. Wan2.1LoRA combines the training technology with few samples, not only reduces hardware requirements, but also improves the accessibility of personalized creation. This approach opens new paths for independent developers and small teams compared to traditional models that require large data sets and high-performance servers.

Project address: https://github.com/ostris/ai-toolkit