Starnodes/quants

FireRed-Image-Edit-1.0 Quantizations (FP8 & NVFP4)

This repository provides quantized versions of the FireRed-Image-Edit-1.0 model, optimized for reduced VRAM usage while maintaining high performance.

We provide weights in FP8 and the cutting-edge NVFP4 (Nvidia FP4) formats, making it easier to run this powerful image editing model on a wider range of hardware. Original source (about the model): https://huggingface.co/FireRedTeam/FireRed-Image-Edit-1.0

Available Versions

File Name	Format	Size	Description
FireRed-Image-Edit-1.0_FP8.safetensors	FP8	~20.4 GB	Standard 8-bit quantization for balanced performance.
FireRed-Image-Edit-1_NVFP4.safetensors	NVFP4	~11.6 GB	Highly compressed 4-bit format optimized for NVIDIA Blackwell/Ada architecture.

About FireRed-Image-Edit

FireRed-Image-Edit is a state-of-the-art model designed for precise and high-quality image editing based on textual instructions. By using these quantized versions, you can:

• Reduce VRAM consumption significantly (especially with the NVFP4 version).
• Speed up inference on compatible hardware.
• Run the model on consumer GPUs that might otherwise lack the memory for the full-precision weights.

Usage

You can use these weights with popular frameworks like ComfyUI or diffusers (ensure you have the necessary libraries installed for FP8/FP4 support).

Requirements

• For NVFP4: Requires latest NVIDIA drivers and hardware support (Ada Lovelace or newer recommended).
• For FP8: Compatible with NVIDIA 30nd/40th series and latest torch/accelerate versions.

Installation

# Example: Download using huggingface-cli
huggingface-cli download Starnodes/quants --local-dir ./models/fire-red-quants