Qwen Image 2.1 Object Remover BBox Turbo

Unofficial mirror of the Qwen Image 2.1 Object Remover BBox Turbo LoRA by prithivMLmods.

What it does

Turbo-optimized LoRA adapter for Qwen-Image-2.1, designed to remove unwanted objects inside user-defined bounding boxes while preserving surrounding textures, lighting, shadows, perspective, and overall image consistency. Supports Turbo and standard workflows.

Model details

• Base model: Qwen/Qwen-Image-2.1

• Adapter: Qwen-Image-2.1-Object-Remover-Bbox-turbo

• Type: LoRA / Adapter

• Status: Experimental

• Inference: Turbo / Standard

• Creator: prithivMLmods

Training specifications

• Dataset: 80 pairs of high-quality images with bounding box annotations and manually manipulated resultant images

• Save precision: BF16

• Learning rate: 1e-4

• Optimizer: AdamW

• Network dimension / rank: 16

• Total steps: 4000

• Trigger prompt: Remove the red highlighted object from the scene

Usage and inference

Load the LoRA adapter with Qwen-Image-2.1 and provide an image with the target object indicated by a clear bounding box.

Trigger prompt: Remove the red highlighted object from the scene

This adapter is optimized for fast Turbo inference while remaining compatible with standard workflows. For Turbo inference, follow the recommended Qwen-Image-2.1 Turbo configuration.

Diffusers example

pip install -U diffusers transformers accelerate

import torch
from diffusers import DiffusionPipeline
from diffusers.utils import load_image

pipe = DiffusionPipeline.from_pretrained(
    "Qwen/Qwen-Image-2.1",
    dtype=torch.bfloat16,
    device_map="cuda",
)
pipe.load_lora_weights(
    "prithivMLmods/Qwen-Image-2.1-Object-Remover-Bbox-turbo"
)
prompt = "Remove the red highlighted object from the scene"
image = load_image("input.png")
result = pipe(image=image, prompt=prompt).images[0]

Comparison summary

Scenario          Base model result                         LoRA result / observation                         Steps
Shadows           May struggle to retain shadows             Better preservation of surrounding shadows        40
Cats              One cat may remain unremoved               Better handling of the marked bounding boxes       40
Objects           May remove unmarked objects; struggles     Experimental behavior with more than two objects   40
Multiple BBox     May struggle with more than one box       Designed for bounding-box object removal            40

Visual comparisons from the original README

1. Shadows

Input

Base model

With LoRA

Base model without LoRA can struggle to retain shadows after removal. Total steps: 40.

2. Cats

Input

Base model

With LoRA

Base model without LoRA can leave one cat unremoved in multi-box cases. Total steps: 40.

3. Objects

Input

Base model

With LoRA

The original comparison notes that unmarked objects may be removed and that more than two objects can be difficult. Total steps: 40.

4. Multiple bounding boxes

Input

Base model

With LoRA

Base model without LoRA can struggle to remove more than one object in a bounding-box scenario. Total steps: 40.

Limitations

Experimental release. Results may vary with complex backgrounds, large or overlapping objects, fine structures and textures, reflections and transparent objects, difficult lighting or perspective, and ambiguous or poorly positioned bounding boxes.

Original source and license

Original repository: prithivMLmods/Qwen-Image-2.1-Object-Remover-Bbox-turbo

License: Qwen Research License Agreement — non-commercial use only. This is an unofficial mirror and is not created or endorsed by the original author. Preserve the original license and attribution notices. Do not sell this model or merges made from it. Built with Qwen.