how to inference with the model without AR during training?

[LaLaLailalai]

Hi @jy0205,

I tried to finetune the dit without AR, so is there any difference when inferencing with this kind of model， especially the i2v task? Thanks!

[feifeiobama]

For I2V there is. Please follow recent papers on video diffusion model (e.g. SVD, Emu Video etc.) to manually add support for I2V generation, since only the AR version can natively support it.

[LaLaLailalai]

For I2V there is. Please follow recent papers on video diffusion model (e.g. SVD, Emu Video etc.) to manually add support for I2V generation, since only the AR version can natively support it.

Thanks! Then what is the difference for t2v generation when loading the model without AR and the one with AR ?