Have you tried training AnyRecon on a smaller I2V base model?

[xiaotaw]

Hi,

WAN I2V 14B is quite a large base model, and needs a lot of memory and compute.

I was wondering if a smaller I2V model could still work well and achieve similar quality results?

Any materials or experiments would be a greatly appreciated.

Thank you!

[yutian10]

Yes, it is definitely possible to achieve comparable results with smaller I2V models.
I have seen people successfully fine-tune models like Wan2.1 1.3B T2V and Wan2.2 5B I2V for similar tasks, and the overall quality can be quite close to the 14B version. The main differences are usually in fine-grained details—smaller models may produce slightly blurrier textures or minor temporal flickering in complex regions. For many practical applications, these trade-offs are acceptable.

[xiaotaw]

So, how long did it take to train 100k steps on the 14B model using 40 frames per step at 512×896?

The original video sequences are partitioned into clips of 40 frames each at the resolution of 512×896.

First, we perform full self-attention fine-tuning for 100k iterations, allowing the model to adapt its internal generative priors to our geometry-controlled input space.