The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

After this PR merged, It will be necessary to implement StableDiffusionMixIn inheritance in SANA pipelines in order to enable tiling and slicing.

@chenjy2003 I think you can implement this: See this signature class with StableDiffusionMixin

class SanaPipeline(DiffusionPipeline, StableDiffusionMixin, SanaLoraLoaderMixin):
    r"""
    Pipeline for text-to-image generation using [Sana](https://huggingface.co/papers/2410.10629).
    """

Ive tested locally with your changes and working.

pipe.enable_vae_slicing()
pipe.enable_vae_tiling()

Dear @lawrence-cj please add this feature to your official SANA pipeline

@elismasilva @a-r-r-o-w I've added the four methods enable_vae_slicing, disable_vae_slicing, enable_vae_tiling, and disable_vae_tiling to SanaPipeline.

@elismasilva @a-r-r-o-w I've added the four methods enable_vae_slicing, disable_vae_slicing, enable_vae_tiling, and disable_vae_tiling to SanaPipeline.

ok i think you forgot to do the same for this pipeline
https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/pag/pipeline_pag_sana.py

@elismasilva @a-r-r-o-w I've added the four methods enable_vae_slicing, disable_vae_slicing, enable_vae_tiling, and disable_vae_tiling to SanaPipeline.

ok i think you forgot to do the same for this pipeline https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/pag/pipeline_pag_sana.py

Thanks for pointing out! @lawrence-cj Could you please double-check whether no other SANA pipelines need to be modified?

We have two pipelines in diffusers. I'm ok with it! Thanks @chenjy2003 for hard work!

Subscribing for it to be merged
NVlabs/Sana#138

what is difference between pipeline_pag_sana.py and pipeline_sana.py

what is difference between pipeline_pag_sana.py and pipeline_sana.py

@lawrence-cj Here is a question for you.

@FurkanGozukara @chenjy2003

It was answered here
NVlabs/Sana#107 (comment)

@FurkanGozukara @chenjy2003

It was answered here NVlabs/Sana#107 (comment)

which one is recommended it doesnt tell much

@FurkanGozukara @chenjy2003

which one is recommended it doesnt tell much

Based on my test results PAG quality is 60-70% lower. The question remains why create PAG pipeline or in which specific scenario it will be useful.

@FurkanGozukara @chenjy2003

which one is recommended it doesnt tell much

Based on my test results PAG quality is 60-70% lower. The question remains why create PAG pipeline or in which specific scenario it will be useful.

thanks a lot

Based on my test results PAG quality is 60-70% lower. The question remains why create PAG pipeline or in which specific scenario it will be useful.

PAG is a function. Anyone can test and try to improve it. We just did the basic work for the community.

PAG is a function. Anyone can test and try to improve it. We just did the basic work for the community.

Makes sense. So it's primarily for developers to improve or enhance it.

this guy Furkan should just be blocked to be honest, creates too much noise and demands a lot from others without thanks

Makes sense. So it's primarily for developers to improve or enhance it.

Correct.

Verified that it works as expected:

import torch
from diffusers import SanaPipeline

pipe = SanaPipeline.from_pretrained(
    "Efficient-Large-Model/Sana_1600M_2Kpx_BF16_diffusers",
    variant="bf16",
    torch_dtype=torch.bfloat16,
)
pipe.to("cuda")
pipe.enable_vae_tiling()

prompt = 'Self-portrait oil painting, a beautiful cyborg with golden hair, 8k'
image = pipe(
    prompt=prompt,
    height=2048,
    width=2048,
    guidance_scale=5.0,
    num_inference_steps=20,
    generator=torch.Generator().manual_seed(42),
)[0]
image[0].save("output.png")

Thank you all! Works beautifully
Something worth mentioning in the docs or model card: the AE's default tile_sample_min_height of 512 creates artifacts on large homogenous areas like a blue sky when generating 4k images. Increasing it to 1024 fixed it for the images I looked at and it still runs without OOM on a 3090/4090.

tile_sample_min_height

can you share your code how do you generate?

can you share your code how do you generate?

notebook @FurkanGozukara, if you ever get famous with your genious YouTube videos, make sure to mention me in your nobel prize speech

can you share your code how do you generate?

notebook @FurkanGozukara, if you ever get famous with your genious YouTube videos, make sure to mention me in your nobel prize speech

thanks a lot. sure if i get famous haha :D

i see you don't use slicing any particular reason?

can you take the unrelated discussion elsewhere?

Thank you all! Works beautifully Something worth mentioning in the docs or model card: the AE's default tile_sample_min_height of 512 creates artifacts on large homogenous areas like a blue sky when generating 4k images. Increasing it to 1024 fixed it for the images I looked at and it still runs without OOM on a 3090/4090.

How much GPU memory is needed for 3090/4090? I tested on A100, requiring 22GB. @geronimi73

How much GPU memory is needed for 3090/4090? I tested on A100, requiring 22GB.

What I meant what was that generating 4096x4096px images with tile_sample_min_height=1024 runs on a GPU with 24GB VRAM like a NVIDIA GefForce RTX 3090 or 4090. Sorry for the confusion.

Oh nevermind. Thanks man @geronimi73

I updated my Gradio APP

4k model now works as low as 8GB GPU VRAM

Diffusers pipeline has amazing improvements

How much GPU memory is needed for 3090/4090? I tested on A100, requiring 22GB.

What I meant what was that generating 4096x4096px images with tile_sample_min_height=1024 runs on a GPU with 24GB VRAM like a NVIDIA GefForce RTX 3090 or 4090. Sorry for the confusion.

My 3060 Ti 8GB runs great, if you enable cpu offload and vae slicing and vae tiling.