HVision-NKU / StoryDiffusion

Accepted as [NeurIPS 2024] Spotlight Presentation Paper
Apache License 2.0
5.92k stars 593 forks source link

How much VRAM do I need to run this on Gradio? #13

Open davizca opened 6 months ago

davizca commented 6 months ago

RTX 3090/4090 can handle this?

And also, are you plan to release the weights on GIthub?

Thanks in advance!

ScotterMonk commented 6 months ago

I, too, am curious about PC requirements. How will the following do? AMD 5800X 64GB, RTX 3080 12GB

smthemex commented 6 months ago

4070 can run

Speedway1 commented 6 months ago

Cannot run onRTX4090 with 24GB. Keep getting CUDA out of memory error.

cryptowooser commented 6 months ago

I'm OOMing as well, is this model just really beefy? Using the defaults, running on linux with Triton etc. installed.

Speedway1 commented 6 months ago

OK we just found a way to run it on a RTX4090 , you need to drop the "number of sample steps" to 35 and the image dimensions should be dropped too. 1040x640 or 640x640 works.And you can only have up to 5 or 6 steps in the description as well. Anything longer blows the ram. Having said that, right now only "RealVision" is working.

deepbeepmeep commented 6 months ago

I have been succesful with a RTX 4090 by enabling VAE slicing and CPU offloading.

Add at line 545 of gradio_app_sdxl_specific_id.py: [TAB]pipe.enable_vae_slicing() [TAB]pipe.enable_model_cpu_offload()

You may need to use an old version of pillow as the one installed by pip may be too recent: pip install pillow==9.5

Speedway1 commented 6 months ago

I have been succesful with a RTX 4090 by enabling VAE slicing and CPU offloading.

Add at line 545 of gradio_app_sdxl_specific_id.py: [TAB]pipe.enable_vae_slicing() [TAB]pipe.enable_model_cpu_offload()

You may need to use an old version of pillow as the one installed by pip may be too recent: pip install pillow==9.5

Thank you that's very helpful!

Z-YuPeng commented 6 months ago

We now add a low GPU Memory cost version, it was tested on a machine with 24GB GPU memory (Tesla A10) and 30GB RAM and is expected to work well with >20 G GPU memory.

python gradio_app_sdxl_specific_id_low_vram.py
maxbizz commented 5 months ago

Any chance we can run this on 3060 12gb vram?

jjhaggar commented 5 months ago

Any way of running this on a Nvidia GeForce RTX 2070 with 8gb vram? Maybe lowering the resolution and number of steps? Please, let us know if somebody gets it to work with similar hardware specs :)

AayushSameerShah commented 5 months ago

I think I should leave this chat having RTX 4050 6GB Laptop GPU 😅