qwen-applications/collectionlora

★ 14PythonAudience · researcherComplexity · 4/5Setup · hard

Mindmap

mindmap
  root((repo))
    What it does
      Merge 50 LoRA adapters
      Single combined adapter
      Faster generation
      Blend two effects
    Technique
      Multi-teacher distillation
      On-policy training
      Emergent composition
    Code Included
      Training scripts
      Inference script
      Batch evaluation
    Built On
      Qwen image model
      Python
    Research
      Zhejiang University
      Alibaba Qwen team
      arXiv 2605.25378

mindmap root((repo)) What it does Merge 50 LoRA adapters Single combined adapter Faster generation Blend two effects Technique Multi-teacher distillation On-policy training Emergent composition Code Included Training scripts Inference script Batch evaluation Built On Qwen image model Python Research Zhejiang University Alibaba Qwen team arXiv 2605.25378

Click or tap to explore — scroll the page freely

Things people build with this

USE CASE 1

Run inference with a single merged adapter that applies multiple image effects without swapping individual LoRA adapters between uses.

USE CASE 2

Train your own combined adapter using the provided training script and a folder of your own teacher LoRA files.

USE CASE 3

Test emergent composition: see the single adapter apply two different visual effects at the same time without any additional training for that combination.

Tech stack

Python

Getting it running

Difficulty · hard Time to first run · 1day+

Model weights were not yet released at time of writing, requires GPU resources and the Qwen image editing base model.

Research code published alongside an academic paper, no explicit license terms are stated in the repository.

In plain English

This repository contains the code for a research paper called CollectionLoRA, published by researchers at Zhejiang University and Alibaba's Qwen team. The project addresses a practical problem in AI image editing: when you want an AI model to handle many different visual effects or styles, you typically need a separate small add-on module (called a LoRA adapter) for each one. Swapping or stacking these modules during use is slow and can cause them to interfere with each other. CollectionLoRA proposes a way to train a single LoRA adapter that contains the knowledge from 50 or more individual effect-specific adapters. The technique, called multi-teacher on-policy distillation, has each specialist adapter act as a teacher during training, and the resulting single adapter learns to replicate all of their behaviors. The paper reports that the combined adapter performs comparably to or better than using each specialist separately, while also being faster to run (requiring only 8 generation steps instead of the typical higher count). An additional property the paper highlights is emergent composition: after training on 50 individual effects, the single adapter can apply two of those effects together at the same time, without any additional training for that combination. The researchers also tested scaling this to 180 teacher adapters compressed into one. The repository includes training scripts and inference code. Running inference requires downloading model weights (a checkpoint folder the README describes in detail) and then running a Python script that applies the loaded adapter to input images. A batch inference script is also provided for evaluating on test sets. Training can be started with a provided shell script and a sample teacher LoRA included in the checkpoint folder. Model weights are listed as not yet released at the time of writing. The code is research-grade and published alongside the arXiv paper (arXiv:2605.25378). It builds on top of Alibaba's Qwen image editing model as its base.

Copy-paste prompts

Prompt 1

I have the CollectionLoRA checkpoint downloaded. Walk me through running the inference script to apply the combined LoRA adapter to my own input images, including the required command-line arguments.

Prompt 2

I want to train a CollectionLoRA model using my own set of teacher LoRA adapters. Walk me through the training shell script and the expected format for the teacher LoRA files.

Prompt 3

How does CollectionLoRA's multi-teacher distillation work? Explain in plain terms how 50 teacher adapters teach a single student adapter during training.

Prompt 4

How does CollectionLoRA achieve emergent composition, where the single adapter can apply two effects at once without being explicitly trained on that combination?

Prompt 5

I want to scale CollectionLoRA to 180 teacher adapters as described in the paper. What changes do I need to make to the training config and what hardware is required?

Open on GitHub → Explain another repo

← qwen-applications on gitmyhub — every repo by this author, as a profile.

Verify against the repo before relying on details.