(This post was authored by hlky and Sayak)

When operating with latent-space diffusion models for high-resolution image and video synthesis, the VAE decoder can consume quite a bit more memory. This makes it hard for the users to run these models on consumer GPUs without going through latency sacrifices and others alike.

For example, with offloading, there is a device transfer overhead, causing delays in the overall inference latency. Tiling is another solution that lets us operate on so-called “tiles” of inputs. However, it can have a negative impact on the quality of the final image.

Therefore, we want to pilot an idea with the community — delegating the decoding process to a remote endpoint.

No data is stored or tracked, and code is open source. We made some changes to huggingface-inference-toolkit and use custom handlers.

This experimental feature is developed by Diffusers 🧨

Table of contents:

Getting started
- Code
- Basic example
- Options
- Generation
- Queueing
Available VAEs
Advantages of using a remote VAE
Provide feedback

Getting started

Below, we cover three use cases where we think this remote VAE inference would be beneficial.

Code

First, we have created a helper method for interacting with Remote VAEs.

Install diffusers from main to run the code. pip install git+https://github.com/huggingface/diffusers@main

Code

from diffusers.utils.remote_utils import remote_decode

Basic example

Here, we show how to use the remote VAE on random tensors.

Code

image = remote_decode(
    endpoint="https://q1bj3bpq6kzilnsu.us-east-1.aws.endpoints.huggingface.cloud/",
    tensor=torch.randn([1, 4, 64, 64], dtype=torch.float16),
    scaling_factor=0.18215,
)

Usage for Flux is slightly different. Flux latents are packed so we need to send the height and width.

Code

image = remote_decode(
    endpoint="https://whhx50ex1aryqvw6.us-east-1.aws.endpoints.huggingface.cloud/",
    tensor=torch.randn([1, 4096, 64], dtype=torch.float16),
    height=1024,
    width=1024,
    scaling_factor=0.3611,
    shift_factor=0.1159,
)

Finally, an example for HunyuanVideo.

Code

video = remote_decode(
    endpoint="https://o7ywnmrahorts457.us-east-1.aws.endpoints.huggingface.cloud/",
    tensor=torch.randn([1, 16, 3, 40, 64], dtype=torch.float16),
    output_type="mp4",
)
with open("video.mp4", "wb") as f:
    f.write(video)

Generation

But we want to use the VAE on an actual pipeline to get an actual image, not random noise. The example below shows how to do it with SD v1.5.

Code

from diffusers import StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained(
    "stable-diffusion-v1-5/stable-diffusion-v1-5",
    torch_dtype=torch.float16,
    variant="fp16",
    vae=None,
).to("cuda")

prompt = "Strawberry ice cream, in a stylish modern glass, coconut, splashing milk cream and honey, in a gradient purple background, fluid motion, dynamic movement, cinematic lighting, Mysterious"

latent = pipe(
    prompt=prompt,
    output_type="latent",
).images
image = remote_decode(
    endpoint="https://q1bj3bpq6kzilnsu.us-east-1.aws.endpoints.huggingface.cloud/",
    tensor=latent,
    scaling_factor=0.18215,
)
image.save("test.jpg")

Here’s another example with Flux.

Code

from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-schnell",
    torch_dtype=torch.bfloat16,
    vae=None,
).to("cuda")

prompt = "Strawberry ice cream, in a stylish modern glass, coconut, splashing milk cream and honey, in a gradient purple background, fluid motion, dynamic movement, cinematic lighting, Mysterious"

latent = pipe(
    prompt=prompt,
    guidance_scale=0.0,
    num_inference_steps=4,
    output_type="latent",
).images
image = remote_decode(
    endpoint="https://whhx50ex1aryqvw6.us-east-1.aws.endpoints.huggingface.cloud/",
    tensor=latent,
    height=1024,
    width=1024,
    scaling_factor=0.3611,
    shift_factor=0.1159,
)
image.save("test.jpg")

Here’s an example with HunyuanVideo.

Code

from diffusers import HunyuanVideoPipeline, HunyuanVideoTransformer3DModel

model_id = "hunyuanvideo-community/HunyuanVideo"
transformer = HunyuanVideoTransformer3DModel.from_pretrained(
    model_id, subfolder="transformer", torch_dtype=torch.bfloat16
)
pipe = HunyuanVideoPipeline.from_pretrained(
    model_id, transformer=transformer, vae=None, torch_dtype=torch.float16
).to("cuda")

latent = pipe(
    prompt="A cat walks on the grass, realistic",
    height=320,
    width=512,
    num_frames=61,
    num_inference_steps=30,
    output_type="latent",
).frames

video = remote_decode(
    endpoint="https://o7ywnmrahorts457.us-east-1.aws.endpoints.huggingface.cloud/",
    tensor=latent,
    output_type="mp4",
)

if isinstance(video, bytes):
    with open("video.mp4", "wb") as f:
        f.write(video)

Queueing

One of the great benefits of using a remote VAE is that we can queue multiple generation requests. While the current latent is being processed for decoding, we can already queue another one. This helps improve concurrency.

Code

import queue
import threading
from IPython.display import display
from diffusers import StableDiffusionPipeline

def decode_worker(q: queue.Queue):
    while True:
        item = q.get()
        if item is None:
            break
        image = remote_decode(
            endpoint="https://q1bj3bpq6kzilnsu.us-east-1.aws.endpoints.huggingface.cloud/",
            tensor=item,
            scaling_factor=0.18215,
        )
        display(image)
        q.task_done()

q = queue.Queue()
thread = threading.Thread(target=decode_worker, args=(q,), daemon=True)
thread.start()

def decode(latent: torch.Tensor):
    q.put(latent)

prompts = [
    "Blueberry ice cream, in a stylish modern glass , ice cubes, nuts, mint leaves, splashing milk cream, in a gradient purple background, fluid motion, dynamic movement, cinematic lighting, Mysterious",
    "Lemonade in a glass, mint leaves, in an aqua and white background, flowers, ice cubes, halo, fluid motion, dynamic movement, soft lighting, digital painting, rule of thirds composition, Art by Greg rutkowski, Coby whitmore",
    "Comic book art, beautiful, vintage, pastel neon colors, extremely detailed pupils, delicate features, light on face, slight smile, Artgerm, Mary Blair, Edmund Dulac, long dark locks, bangs, glowing, fashionable style, fairytale ambience, hot pink.",
    "Masterpiece, vanilla cone ice cream garnished with chocolate syrup, crushed nuts, choco flakes, in a brown background, gold, cinematic lighting, Art by WLOP",
    "A bowl of milk, falling cornflakes, berries, blueberries, in a white background, soft lighting, intricate details, rule of thirds, octane render, volumetric lighting",
    "Cold Coffee with cream, crushed almonds, in a glass, choco flakes, ice cubes, wet, in a wooden background, cinematic lighting, hyper realistic painting, art by Carne Griffiths, octane render, volumetric lighting, fluid motion, dynamic movement, muted colors,",
]

pipe = StableDiffusionPipeline.from_pretrained(
    "Lykon/dreamshaper-8",
    torch_dtype=torch.float16,
    vae=None,
).to("cuda")

pipe.unet = pipe.unet.to(memory_format=torch.channels_last)
pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)

_ = pipe(
    prompt=prompts[0],
    output_type="latent",
)

for prompt in prompts:
    latent = pipe(
        prompt=prompt,
        output_type="latent",
    ).images
    decode(latent)

q.put(None)
thread.join()

Available VAEs

Endpoint

Model

Stable Diffusion v1

https://q1bj3bpq6kzilnsu.us-east-1.aws.endpoints.huggingface.cloud

stabilityai/sd-vae-ft-mse

Stable Diffusion XL

https://x2dmsqunjd6k9prw.us-east-1.aws.endpoints.huggingface.cloud

madebyollin/sdxl-vae-fp16-fix

Flux

https://whhx50ex1aryqvw6.us-east-1.aws.endpoints.huggingface.cloud

black-forest-labs/FLUX.1-schnell

HunyuanVideo

https://o7ywnmrahorts457.us-east-1.aws.endpoints.huggingface.cloud

hunyuanvideo-community/HunyuanVideo

Advantages of using a remote VAE

These tables demonstrate the VRAM requirements with different GPUs. Memory usage % determines whether users of a certain GPU will need to offload. Offload times vary with CPU, RAM and HDD/NVMe. Tiled decoding increases inference time.

SD v1.5

GPU

Resolution

Time (seconds)

Memory (%)

Tiled Time (secs)

Tiled Memory (%)

NVIDIA GeForce RTX 4090

512x512

0.031

5.60%

0.031 (0%)

5.60%

NVIDIA GeForce RTX 4090

1024x1024

0.148

20.00%

0.301 (+103%)

5.60%

NVIDIA GeForce RTX 4080

512x512

0.05

8.40%

0.050 (0%)

8.40%

NVIDIA GeForce RTX 4080

1024x1024

0.224

30.00%

0.356 (+59%)

8.40%

NVIDIA GeForce RTX 4070 Ti

512x512

0.066

11.30%

0.066 (0%)

11.30%

NVIDIA GeForce RTX 4070 Ti

1024x1024

0.284

40.50%

0.454 (+60%)

11.40%

NVIDIA GeForce RTX 3090

512x512

0.062

5.20%

0.062 (0%)

5.20%

NVIDIA GeForce RTX 3090

1024x1024

0.253

18.50%

0.464 (+83%)

5.20%

NVIDIA GeForce RTX 3080

512x512

0.07

12.80%

0.070 (0%)

12.80%

NVIDIA GeForce RTX 3080

1024x1024

0.286

45.30%

0.466 (+63%)

12.90%

NVIDIA GeForce RTX 3070

512x512

0.102

15.90%

0.102 (0%)

15.90%

NVIDIA GeForce RTX 3070

1024x1024

0.421

56.30%

0.746 (+77%)

16.00%

SDXL

GPU

Resolution

Time (seconds)

Memory Consumed (%)

Tiled Time (seconds)

Tiled Memory (%)

NVIDIA GeForce RTX 4090

512x512

0.057

10.00%

0.057 (0%)

10.00%

NVIDIA GeForce RTX 4090

1024x1024

0.256

35.50%

0.257 (+0.4%)

35.50%

NVIDIA GeForce RTX 4080

512x512

0.092

15.00%

0.092 (0%)

15.00%

NVIDIA GeForce RTX 4080

1024x1024

0.406

53.30%

0.406 (0%)

53.30%

NVIDIA GeForce RTX 4070 Ti

512x512

0.121

20.20%

0.120 (-0.8%)

20.20%

NVIDIA GeForce RTX 4070 Ti

1024x1024

0.519

72.00%

0.519 (0%)

72.00%

NVIDIA GeForce RTX 3090

512x512

0.107

10.50%

0.107 (0%)

10.50%

NVIDIA GeForce RTX 3090

1024x1024

0.459

38.00%

0.460 (+0.2%)

38.00%

NVIDIA GeForce RTX 3080

512x512

0.121

25.60%

0.121 (0%)

25.60%

NVIDIA GeForce RTX 3080

1024x1024

0.524

93.00%

0.524 (0%)

93.00%

NVIDIA GeForce RTX 3070

512x512

0.183

31.80%

0.183 (0%)

31.80%

NVIDIA GeForce RTX 3070

1024x1024

0.794

96.40%

0.794 (0%)

96.40%

Provide feedback

If you like the idea and feature, please help us with your feedback on how we can make this better and whether you’d be interested in having this kind of feature more natively integrated into the Hugging Face ecosystem. If this pilot goes well, we plan on creating optimized VAE endpoints for more models, including the ones that can generate high-resolution videos!

Steps:

Open an issue on Diffusers through this link.
Answer the questions and provide any extra info you want.
Hit submit!

Remote VAEs for decoding with Inference Endpoints 🤗

Getting started

Code

Basic example

Generation

Queueing

Available VAEs

Advantages of using a remote VAE

Provide feedback

Steps: