FireRed-Image-Edit-1.0-FP8

FireRed-Image-Edit-1.0-FP8 is an FP8-compressed version of FireRed-Image-Edit-1.0 from FireRed Team, delivering the same state-of-the-art instruction-based diffusion transformer image editing capabilities with substantially reduced memory footprint and enhanced inference efficiency. Built upon the original 1.6B-sample corpus refined into over 100M high-quality text-to-image and editing pairs through large-scale cleaning, stratification, auto-labeling, and dual-stage semantic filtering, this FP8 release preserves the model’s multi-stage training pipeline of pre-training, supervised fine-tuning, and reinforcement learning, including innovations such as the Multi-Condition Aware Bucket Sampler for variable resolutions, Stochastic Instruction Alignment, Asymmetric Gradient Optimization for stable DPO training, DiffusionNFT with layout-OCR reward modeling for precise text editing, and differentiable Consistency Loss for strong identity preservation. Optimized for efficient deployment, the FP8 compression significantly lowers VRAM requirements while maintaining high semantic alignment, visual fidelity, and identity consistency across diverse editing tasks such as photo restoration, multi-image virtual try-on, style transfer with accurate text rendering, object insertion and modification, and layout-aware editing, with seamless integration into ComfyUI workflows for research and production-grade pipelines.

Download FP8

• FireRed-Image-Edit-1.0 @ FP8 - Download

Quick Start with Diffusers 🧨

Install the required packages

transformers # - transformers@v4.57.6
torch        # - torch@v2.9.1+cu128
diffusers    # - diffusers@v0.37.0.dev0
accelerate   # - accelerate@v1.12.0

FireRed-Image-Edit-1.0-FP8 [Demo]

import torch
from diffusers import QwenImageEditPipeline  # or your compatible pipeline

model_path = "./FireRed-Image-Edit-1.0-FP8"

pipe = QwenImageEditPipeline.from_pretrained(
    model_path,
    torch_dtype=torch.bfloat16
)

pipe.to("cuda")

# Example usage
# outputs = pipe(image=input_image, prompt="Your edit prompt")