fp8 @ - models

Qwen-Image-Edit-AIO-fp8

Qwen-Image-Edit-AIO-fp8 is an FP8-compressed Transformers-only release built from Qwen-Image-Edit-2511 and Qwen-Image-Edit-2509. This repository provides FP8 (W8A8 · F8_E4M3) compressed transformer weights compatible with both Transformers and Diffusers pipelines, optimized for lower VRAM usage and higher throughput while maintaining strong editing fidelity.

Diffusers Usage

Qwen-Image-Edit-2509-fp8

import torch
from diffusers.models import QwenImageTransformer2DModel
from diffusers import QwenImageEditPlusPipeline
from diffusers.utils import load_image

# Load transformer (2509 version)
transformer = QwenImageTransformer2DModel.from_pretrained(
    "prithivMLmods/FireRed-Image-Edit-1.0-fp8",
    subfolder="Qwen-Image-Edit-2509-fp8/transformer",
    torch_dtype=torch.bfloat16
)

# Load pipeline
pipeline = QwenImageEditPlusPipeline.from_pretrained(
    "Qwen/Qwen-Image-Edit-2509",
    transformer=transformer,
    torch_dtype=torch.bfloat16
)

pipeline.to("cuda")

# Load image
image1 = load_image("grumpycat.png")
prompt = "turn the cat into an orange cat"

# Inference inputs
inputs = {
    "image": [image1],
    "prompt": prompt,
    "generator": torch.manual_seed(42),
    "true_cfg_scale": 1.0,
    "negative_prompt": " ",
    "num_inference_steps": 40,
    "guidance_scale": 1.0,
    "num_images_per_prompt": 1,
}

# Run pipeline
output = pipeline(**inputs)
output_image = output.images[0]
output_image.save("output_image_2509.png")

Qwen-Image-Edit-2511-fp8

import torch
from diffusers.models import QwenImageTransformer2DModel
from diffusers import QwenImageEditPlusPipeline
from diffusers.utils import load_image

# Load transformer (2511 version)
transformer = QwenImageTransformer2DModel.from_pretrained(
    "prithivMLmods/FireRed-Image-Edit-1.0-fp8",
    subfolder="Qwen-Image-Edit-2511-fp8/transformer",
    torch_dtype=torch.bfloat16
)

# Load pipeline
pipeline = QwenImageEditPlusPipeline.from_pretrained(
    "Qwen/Qwen-Image-Edit-2511",
    transformer=transformer,
    torch_dtype=torch.bfloat16
)

pipeline.to("cuda")

# Load image
image1 = load_image("grumpycat.png")
prompt = "turn the cat into an orange cat"

# Inference inputs
inputs = {
    "image": [image1],
    "prompt": prompt,
    "generator": torch.manual_seed(42),
    "true_cfg_scale": 1.0,
    "negative_prompt": " ",
    "num_inference_steps": 40,
    "guidance_scale": 1.0,
    "num_images_per_prompt": 1,
}

# Run pipeline
output = pipeline(**inputs)
output_image = output.images[0]
output_image.save("output_image_2511.png")

---

About the Base Models

Qwen-Image-Edit-2511 is an advanced iteration over Qwen-Image-Edit-2509, a production-grade 20B-parameter image editing model developed by Alibaba's Qwen team.

It is built on the Qwen-Image MMDiT architecture with VL-Qwen2.5 integration and VAE encoding for high-fidelity real-world editing workflows.

Architecture Overview

DiT / MMDiT Block	U-Net Architecture
Cross-Attention Mechanism	Variational Autoencoder (VAE)

• Qwen-Image MMDiT backbone
• VL-Qwen2.5 multimodal integration
• Latent VAE encoding
• Diffusion-based editing pipeline
• Compatible with Diffusers and ComfyUI workflows

What Makes 2511 Stronger Than 2509

Multi-Person & Identity Consistency

• Reduced identity swaps in group photos
• Improved pose consistency
• Stable multi-subject generation

High-Fidelity Iterative Editing

• Strong identity preservation across multiple edits
• Reduced image drift
• Reliable prompt alignment

Dual-Mode Editing Control

• Appearance editing for localized modifications
• Semantic editing for global transformations

Precise Text Handling

• Natural typography rendering
• Accurate on-image text replacement
• Reduced distortions in signage or UI text

Enhanced Structural Reasoning

• Better geometric alignment
• Accurate object replacement
• Material-aware transformations

Industrial & Commercial Use

• Product design material swaps
• Clean geometry preservation
• Suitable for e-commerce and marketing workflows
• Reliable outputs for creative production pipelines

What This FP8 AIO Version Provides

Qwen-Image-Edit-AIO-fp8 introduces:

• FP8 (W8A8 · F8_E4M3) transformer compression
• Reduced VRAM requirements
• Higher inference throughput
• Transformers-native loading
• Diffusers-compatible transformer weights
• Optimized for Hopper-class and compatible GPUs

This release focuses strictly on compressed transformer weights while preserving original editing capabilities.

Intended Workflows

• ComfyUI pipelines
• Diffusers-based applications
• Production-grade editing systems
• Batch commercial pipelines
• E-commerce product editing
• Marketing creative generation

Limitations

• FP8 acceleration requires compatible GPU architectures.
• Extremely fine-grained edge cases may show minor precision differences compared to full BF16.
• Users are responsible for lawful and ethical deployment.