Pruna AI has released the P-Q-Image-2.1-Remove-Object set of LoRA adapters for the Qwen-Image-2.1 model: they remove an object from a photo along with its shadow and reflection, without shifting the remaining pixels. Distilled variants handle the task in a few generation steps instead of forty and operate without CFG, while the weights are open and free on Hugging Face.

image
image

What happened

Pruna AI has published the P-Q-Image-2.1-Remove-Object set of LoRA adapters for the Qwen-Image-2.1 image generation model on Hugging Face. The adapters remove objects from photos along with their shadows and reflections, without shifting the remaining pixels; shadows and reflections are removed even outside the marked area. Three variants have been released: distilled versions for 6 and 8 steps, and a base version for 40 steps. The fast variants operate without CFG at true_cfg_scale=1.0 and an adapter strength of 0.7, and require the installed distillation LoRA PrunaAI/P-Q-Image-2.1; the 6-step variant is recommended by default. The object is marked with a rough red outline, which is the recommended mode, or simply named in a text instruction. A separate release includes a repository for ComfyUI with ready-made workflows and a custom node.

Context

A LoRA adapter is a compact extension of a base model that adds new behavior without affecting the weights of Qwen-Image-2.1 itself. The main feature of this release is the application of distillation: usually, this is used to speed up the full generative pipeline, but here 40 steps are compressed to 6 or 8 for a single narrow inpainting task, i.e., editing part of an image. The fast variants are built as a stack of two adapters: the task-specific LoRA is placed on top of the distillation LoRA, and the composability of the adapters itself becomes a mechanism for assembling the pipeline. The rejection of CFG makes each step approximately twice as cheap, so the total computational savings per operation are significantly greater than the formal compression from 40 to 6 steps. Essentially, this is an engineering artifact rather than a scientific paper—a public tool designed for practical application.

Why this matters for the industry

For the industry, the release shows that distillation turns individual operations into cheap building blocks, not just speeding up the entire model. The combination of a rough mask, text instruction, and red outline lowers the barrier to entry for automating photo cleanup in production, while open weights allow teams with their own GPU infrastructure to integrate object removal into their own pipelines without an API intermediary. Openness makes the release verifiable: anyone can download the adapters and reproduce the result, so weaknesses will be revealed by the community quite quickly. If the adapters prove to be stable, a wave of similar narrow-task distilled LoRAs for other point operations and base models is likely. In such a scenario, the value of products shifts from the function itself to UX, domain data, and distribution, and thin wrappers over cloud APIs for simple editing operations lose their meaning.

Why this matters for users

If you work with ComfyUI or diffusers, you can download the weights for free from Hugging Face and run object removal locally without calling an API. For ComfyUI, there are ready-made workflows and a custom node, and you can check the result without a local installation in the online demo at dashboard.pruna.ai/playground. When deploying, it is advisable to set the 6-step variant as the default, not forgetting the base distillation LoRA PrunaAI/P-Q-Image-2.1. A realistic first scenario is point cleanup of individual photos in content pipelines; for images below 1 MP, one should not expect a perfectly restored texture.

What is currently unknown / limitations

Limitations are declared by the authors in advance: with a rough mask, the model may remove a neighboring object, and the filled area sometimes turns out smoother than the surroundings, with a slight mesh-like artifact structure. There is no quantitative evaluation methodology in the release: metrics for preservation, latency measurements, and VRAM requirements are absent, so qualitative caveats do not yet replace systematic measurement of failure rates. The fast variants require installing two adapters at once, the task-specific one on top of the distillation LoRA PrunaAI/P-Q-Image-2.1 with a strength of 0.7, which complicates the pipeline configuration. Speculations about a future wave of narrow-task adapters and the standardization of the rough mask and text instruction combination are interpretations of a trend, not verified facts.

Sources

Author

Look at AI, editorial team