One Small Step in Latent, One Giant Leap for Pixels:
Fast Latent Upscale Adapter for Your Diffusion Models

Aleksandr Razin^* · Danil Kazantsev^* · Ilya Makarov

This repository contains the official implementation of the paper "One Small Step in Latent, One Giant Leap for Pixels: Fast Latent Upscale Adapter for Your Diffusion Models".

We present the Latent Upscaler Adapter (LUA), a lightweight module that performs super-resolution directly on the generator's latent code before the final VAE decoding step. LUA integrates as a drop-in component, requiring no modifications to the base model or additional diffusion stages. It enables high-resolution synthesis through a single feed-forward pass in latent space, achieving comparable perceptual quality to pixel-space methods while reducing decoding and upscaling time.

Installation

git clone https://github.com/vaskers5/LUA.git
cd LUA
pip install -r requirements.txt

Quick Start

LUA weights are hosted on HuggingFace and downloaded automatically on first use.

Python API

import torch
from diffusers import FluxPipeline
from lua import load_model, upscale_latent

# Load models
pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16)
pipe.to("cuda")
pipe.vae.enable_tiling()

lua_model = load_model(device="cuda")  # auto-downloads weights from HF

# Generate base latent at 1024x1024
result = pipe("a cat astronaut", output_type="latent", width=1024, height=1024)

# Unpack to VAE space
latent = pipe._unpack_latents(result.images, 1024, 1024, pipe.vae_scale_factor)
latent = (latent / pipe.vae.config.scaling_factor) + pipe.vae.config.shift_factor

# Upscale x2 (1024 -> 2048) or x4 (1024 -> 4096)
upscaled = upscale_latent(lua_model, latent, head="x2")

# Decode to image
image = pipe.vae.decode(upscaled.to(torch.bfloat16), return_dict=False)[0]
image = pipe.image_processor.postprocess(image, output_type="pil")[0]
image.save("output_2k.png")

CLI Inference

# 2K image (1024 -> 2048)
python inference.py --prompt "a mountain landscape, cinematic" --head x2

# 4K image (1024 -> 4096)
python inference.py --prompt "a mountain landscape, cinematic" --head x4 --output landscape_4k.png

# Use a local checkpoint
python inference.py --prompt "hello" --weights ./my_checkpoint.pth --head x2

Gradio Demo

Interactive demo with side-by-side comparison against direct FLUX generation:

python gradio_demo.py

The demo compares LUA path (FLUX@1024 + LUA upscale) vs Direct path (FLUX@target) at the same output resolution, with interactive magnifying loupes and timing breakdowns.

You can configure the FLUX model via environment variables:

FLUX_MODEL_ID="black-forest-labs/FLUX.1-dev" python gradio_demo.py

Model Details

Architecture	SwinIR-based transformer with multi-head upsampling
Parameters	~250M
Input	16-channel VAE latent (FLUX latent space)
Heads	x2 (2x upscaling), x4 (4x upscaling)
Output	Upscaled 16-channel VAE latent

LUA operates entirely in the latent space — it upscales the latent code before the VAE decoder, which means the expensive VAE decode only happens once at the target resolution.

Training

Training code will be released soon.

Citation

@article{razin2024lua,
  title={One Small Step in Latent, One Giant Leap for Pixels: Fast Latent Upscale Adapter for Your Diffusion Models},
  author={Razin, Aleksandr and Kazantsev, Danil and Makarov, Ilya},
  journal={arXiv preprint arXiv:2511.10629},
  year={2024}
}

License

This project is licensed under the Apache License 2.0 — see the LICENSE file for details.