blanchon/nitro_e_512

Nitro-E 512px - Diffusers Integration

This is the Nitro-E 512px text-to-image diffusion model in diffusers format.

Model Description

Nitro-E is a family of text-to-image diffusion models focused on highly efficient training. With just 304M parameters, Nitro-E is designed to be resource-friendly for both training and inference.

Key Features:

• 304M parameters
• Efficient training: 1.5 days on 8x AMD Instinct MI300X GPUs
• High throughput: 18.8 samples/second on single MI300X
• Consumer GPU support: 0.16s per 512px image on Strix Halo iGPU

Model Variant

This is the 512px variant, optimized for generating 512x512 images.

Note: This variant uses Alternating Subregion Attention (ASA) for efficiency.

Original Model

This model is based on amd/Nitro-E and has been converted to the diffusers format for easier integration and use.

Usage

import torch
from diffusers import NitroEPipeline

# Load pipeline
pipe = NitroEPipeline.from_pretrained("blanchon/nitro_e_512", torch_dtype=torch.bfloat16)
pipe = pipe.to("cuda")

# Generate 512x512 image
prompt = "A hot air balloon in the shape of a heart grand canyon"
image = pipe(
    prompt=prompt,
    width=512,
    height=512,
    num_inference_steps=20,
    guidance_scale=4.5,
).images[0]

image.save("output.png")

Technical Details

Architecture

• Type: E-MMDiT (Efficient Multi-scale Masked Diffusion Transformer)
• Attention: Alternating Subregion Attention (ASA)
• Text Encoder: Llama-3.2-1B
• VAE: DC-AE-f32c32 from MIT-Han-Lab
• Scheduler: Flow Matching with Euler Discrete Scheduler
• Sample Size: 16 (latent space)

Training

• Dataset: ~25M images (real + synthetic)
• Duration: 1.5 days on 8x AMD Instinct MI300X GPUs
• Training Details: See Nitro-E Technical Report

Citation

If you use this model, please cite:

@article{nitro-e-2025,
  title={Nitro-E: Efficient Training of Diffusion Models},
  author={AMD AI Group},
  journal={arXiv preprint arXiv:2510.27135},
  year={2025}
}

License

Copyright (c) 2025 Advanced Micro Devices, Inc. All Rights Reserved.

Licensed under the MIT License. See the LICENSE for details.

Related Projects

• Nitro-T: Efficient Training of diffusion models
• Nitro-1: One-step distillation of diffusion models
• Original Nitro-E Repository
• AMD Nitro-E on HuggingFace