ControlNeXt-SVD-v2 out-of-memory on 24GB vram

[sumitmamoria]

What is the minimum VRAM requirement to run ControlNeXt-SVD-v2 ?

[sumitmamoria]

Is it possible to use dual GPU setup?

[Pbihao]

In the v2 version, we implemented a significant change with larger resolution and extended batch frames to achieve better performance. To reduce GPU consumption, you can follow these steps:

1. Use FP16 or BF16 for inference.
2. Reduce the resolution, such as to 768x512.
3. Decrease the number of batch_frames.

However, such operations may influce the final performance

[sumitmamoria]

In the v2 version, we implemented a significant change with larger resolution and extended batch frames to achieve better performance. To reduce GPU consumption, you can follow these steps:

1. Use FP16 or BF16 for inference.
2. Reduce the resolution, such as to 768x512.
3. Decrease the number of batch_frames.

However, such operations may influce the final performance

Thanks. Will try these changes. Is there a way to support dual GPUs?

[Rodmuzik]

Dual GPU +1

[nitinmukesh]

In the v2 version, we implemented a significant change with larger resolution and extended batch frames to achieve better performance. To reduce GPU consumption, you can follow these steps:

1. Use FP16 or BF16 for inference.
2. Reduce the resolution, such as to 768x512.
3. Decrease the number of batch_frames.

Hello @Pbihao

I am successfully able to include point 1, 2, & 3.

Also I want to use this fp16 version instead of unet.bin. Is it possible. I did tried specifying the below file with --unet_path but it's not working.

https://huggingface.co/Kijai/ControlNeXt-SVD-V2-Comfy/tree/main
controlnext-svd_v2-unet-fp16_converted.safetensors

I am not a coder but if you suggest the code changes I will include them.