LUA-FLUX: Latent Upscale Adapter for FLUX

Official weights for LUA (Latent Upscale Adapter) from the paper "One Small Step in Latent, One Giant Leap for Pixels".

LUA is a lightweight (~250M params) SwinIR-based module that upscales diffusion model latents before VAE decoding. It enables high-resolution image generation (2K/4K) from a 1024x1024 base generation, significantly faster than generating at full resolution directly.

Key Features

• Drop-in adapter: works with any FLUX model, no fine-tuning needed
• Dual-head: x2 (1024->2048) and x4 (1024->4096) upscaling from a single model
• Fast: operates in latent space (~250ms for x2, ~900ms for x4 on A100)
• Quality: comparable perceptual quality to direct high-resolution generation

Quick Start

from lua import load_model, upscale_latent

# Auto-downloads this checkpoint
lua_model = load_model(device="cuda")

# Upscale a FLUX VAE latent (B, 16, H, W)
upscaled = upscale_latent(lua_model, latent, head="x2")  # or "x4"

See the GitHub repo for full usage examples including CLI and Gradio demo.

Model Details

Architecture	SwinIR with multi-head upsampling
Parameters	~250M
Input	16-channel VAE latent (FLUX latent space)
Heads	x2 (2x upscaling), x4 (4x upscaling)
Training	Trained on high-quality image pairs in FLUX latent space
Precision	float32 (recommended), bfloat16 compatible

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

Apache 2.0