Views
No views yet
| VAE Variant | Decoding Speed | Quality |
|---|---|---|
| DC-AE (Standard) | 1.0× | Reference |
| DC-AE-Lite | 1.8× | Similar |
1import torch
2from diffusers import NitroEPipeline
3
4# Load the lite variant
5pipe = NitroEPipeline.from_pretrained(
6 "blanchon/nitro_e_512_lite",
7 torch_dtype=torch.bfloat16
8)
9pipe.to("cuda")
10
11# Generate image (1.8x faster decoding!)
12prompt = "A hot air balloon in the shape of a heart grand canyon"
13image = pipe(
14 prompt=prompt,
15 width=512,
16 height=512,
17 num_inference_steps=20,
18 guidance_scale=4.5,
19).images[0]
20
21image.save("output.png")1@article{nitro-e-2025,
2 title={Nitro-E: Efficient Training of Diffusion Models},
3 author={AMD AI Group},
4 journal={arXiv preprint arXiv:2510.27135},
5 year={2025}
6}