Qwen-Image-Edit-2511-bnb-nf4

4-bit NF4 quantized version of Qwen-Image-Edit-2511 using BitsAndBytes.

This quantized model significantly reduces VRAM requirements, making it accessible on consumer GPUs like RTX 3090/4080/4090.

Sample Outputs

Single Image Editing

Multi-Image Editing

Combine multiple images into one coherent scene:

Style Transfer

Features

• Reduced VRAM: ~17GB vs ~40GB (original BF16)
• Same capabilities: Single/multi-image editing, style transfer, character consistency
• BitsAndBytes NF4: High-quality 4-bit quantization

Requirements

pip install torch diffusers transformers accelerate bitsandbytes

Usage

import torch
from PIL import Image
from diffusers import QwenImageEditPlusPipeline

# Load quantized model
pipe = QwenImageEditPlusPipeline.from_pretrained(
    "seochan99/Qwen-Image-Edit-2511-bnb-nf4",
    torch_dtype=torch.bfloat16,
)
pipe.to("cuda")

# Single image editing
image = Image.open("input.png")
result = pipe(
    image=[image],
    prompt="Turn this into anime style",
    true_cfg_scale=4.0,
    negative_prompt=" ",
    num_inference_steps=50,
).images[0]
result.save("output.png")

Multi-Image Editing

Combine two images into one coherent scene:

img1 = Image.open("person.png")
img2 = Image.open("background.png")

result = pipe(
    image=[img1, img2],
    prompt="Person standing in the forest, natural lighting",
    true_cfg_scale=4.0,
    negative_prompt=" ",
    num_inference_steps=50,
).images[0]

VRAM Comparison

Tested on NVIDIA RTX 4090.

Quantization Details

• Method: BitsAndBytes NF4 (4-bit Normal Float)
• Compute dtype: bfloat16
• Components quantized: text_encoder, transformer

Model Info

• Base model: Qwen/Qwen-Image-Edit-2511
• Paper: arXiv:2508.02324
• Original repo: QwenLM/Qwen-Image

License

Apache 2.0 (same as base model)