Same tactics as used by alibaba wan team

[OrangeUnknownCat]

hi,
You opened weights saying these weights are same as closed source weights, and yet there are limitations for example i generated a video with this just single text prompt in API and it worked:

"follow text instructions from first frame"

and closed source model generated a video correctly:

Hailuo_Video_follow.text.instructions.from._541712491691143176.mp4

now look at the output from open weights with same prompt but i even tried different and it never worked:

here the result from both open weights:

https://github.com/user-attachments/assets/20d2975a-4f30-4541-8e8e-0dd126ebaf96
https://github.com/user-attachments/assets/acec3afa-6bdb-4750-aa4a-957cd0ac60a1

one with ref2vid and other fl2vid.

And i must say what is this? i mean it contains a large 32b vl encoder and it still fails to understand simple tasks? i am not saying that the models are bad in fact they are the best one so far as opensource, but selling with fake cherry on top of it is not right.

[MeiYi-dev]

hi, You opened weights saying these weights are same as closed source weights, and yet there are limitations for example i generated a video with this just single text prompt in API and it worked:

"follow text instructions from first frame"

and closed source model generated a video correctly:

Hailuo_Video_follow.text.instructions.from._541712491691143176.mp4
now look at the output from open weights with same prompt but i even tried different and it never worked:

here the result from both open weights:

https://github.com/user-attachments/assets/20d2975a-4f30-4541-8e8e-0dd126ebaf96 https://github.com/user-attachments/assets/acec3afa-6bdb-4750-aa4a-957cd0ac60a1

one with ref2vid and other fl2vid.

And i must say what is this? i mean it contains a large 32b vl encoder and it still fails to understand simple tasks? i am not saying that the models are bad in fact they are the best one so far as opensource, but selling with fake cherry on top of it is not right.

I am not from the Minimax team, but theeir API version uses a prompt enhancer. No model can understand simple words literally, a prompt enhancer is needed for every model. That's how the DiT arch works. They literally state it on their HF page. 32B for text encoding because video models need complex understanding which can only happen in large LLMs. Every multi-shot model out there use large LLMs for prompt encoding.