Regarding the Compute Resources

[amandpkr]

Hi Author,

Congratulations on the excellent work! I wanted to ask how much GPU resources and compute time were required to run your pipeline.

[shim0114]

Thank you for your interest in our work.

For Latte and CogVideoX experiments, we use an A100 GPU with 40GB.
For Wan 2.1 experiments, we use four H100 GPUs with 80 GB.

[amandpkr]

Thank you for your quick repsone @shim0114.

I am kind of confused why do WAN2.1 (1.3B) requires 4 H100 gpus?

[shim0114]

I'm sorry for the confusion.

The 4 H100 GPUs are intended for the Wan 2.1-14B experiments.

We also used 4 H100s for the Wan 2.1-1.3B experiments, which may be overkill.
As our experiments mainly focus on the 14B model, we did not investigate the minimum hardware specifications for the 1.3B model.

[amandpkr]

Thank you @shim0114 for the quick reply.

I am generating images using WAN 2.1-1.3B.
With the original WAN, producing 81 frames takes 1 minute 55 seconds, whereas WAN + DLBS takes 49 minutes 20 seconds for the same number of frames.

I am running everything on a single H100.

Do you think such a large difference in runtime is expected?

[shim0114]

I appreciate you taking the time to follow up.

If DLBS K=2, B=4 is used and # of frames = 81, then it is expected.
We generate 8 times as many latents during generation and perform decoding and reward evaluation for each latent.

The runtime difference may decrease if you set # of frames to 33, for example, because decoding may be the bottleneck in our method.

[amandpkr]

Thank you for the clarification. However, I still don’t fully understand why the 14B model requires four GPUs, especially since the vanilla WAN 2.1 (14B) can generate images on a single H100.

[shim0114]

Thank you for your additional follow up. One possible reason is the memory overhead due to the reward models. Memory usage may have room for optimization.

[amandpkr]

Hey @shim0114,
Thank you for your quick response.
Could you please let me know how long it takes to generate 81 frames on 4×H100 GPUs using WAN 2.1 14B model? Since I don’t have access to four H100s, I’m trying to understand the expected generation time.

[shim0114]

Hello @VIROBO-15,

Thank you for your follow-up, and apologies for the late reply.

In our paper, we did not conduct experiments generating 81 frames using the WAN 2.1 14B model on 4×H100 GPUs.
Instead, we provide the inference costs for the following settings as a reference:

Model	Frames	Hardware	Inference Cost
WAN 2.1 1.3B	81	4×A100	1h 5m 44s
WAN 2.1 14B	33	4×H100	24m 27s

I hope this data gives you a rough idea of the expected generation time.