Nitro-E 512px Lite - Fast Decoding Variant

This is the Nitro-E 512px text-to-image diffusion model with DC-AE-Lite for faster image decoding.

Key Features

🚀 1.8× Faster Decoding: Uses DC-AE-Lite instead of standard DC-AE
🎯 Same Quality: Similar reconstruction quality to standard DC-AE
⚡ Drop-in Compatible: Uses the same Nitro-E transformer weights
💾 Memory Efficient: Smaller decoder footprint

Performance Comparison

VAE Variant	Decoding Speed	Quality
DC-AE (Standard)	1.0×	Reference
DC-AE-Lite	1.8×	Similar

This makes Nitro-E even faster for real-time applications!

Model Details

Transformer: Nitro-E 512px (304M parameters)
VAE: DC-AE-Lite-f32c32 (faster decoder)
Text Encoder: Llama-3.2-1B
Scheduler: Flow Matching with Euler Discrete
Attention: Alternating Subregion Attention (ASA)

Usage

import torch
from diffusers import NitroEPipeline

# Load the lite variant
pipe = NitroEPipeline.from_pretrained(
    "blanchon/nitro_e_512_lite",
    torch_dtype=torch.bfloat16
)
pipe.to("cuda")

# Generate image (1.8x faster decoding!)
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")

When to Use This Variant

Use DC-AE-Lite (this model) when:

You need faster inference
Running real-time applications
Batch processing many images
Decoding is your bottleneck

Use standard DC-AE when:

You need absolute best reconstruction quality
Decoding speed is not critical

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-Lite-f32c32 (1.8× faster decoding)
Scheduler: Flow Matching with Euler Discrete Scheduler
Latent Size: 16×16 for 512×512 images

Recommended Settings

Steps: 20 (good quality/speed trade-off)
Guidance Scale: 4.5 (balanced)
Resolution: 512×512 (optimized)

Citation

@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

Licensed under the MIT License.

blanchon/nitro_e_512_lite