internlm/ETCHR-FLUX.2-klein-9B

ETCHR-FLUX.2-klein-9B

📖<a href="https://arxiv.org/abs/2605.23897">Paper</a> | 🏠<a href="https://github.com/InternLM/ETCHR">Homepage</a > | 🤗<a href="https://huggingface.co/internlm/ETCHR-FLUX.2-klein-9B">ETCHR-FLUX.2-klein-9B Model</a > | 🤗<a href="https://huggingface.co/datasets/BeichenZhang/ETCHR-SFT-400K">ETCHR SFT-400K Dataset</a > | 🤗<a href="https://huggingface.co/datasets/internlm/ETCHR-GRPO-10K">ETCHR GRPO-10K Dataset</a > | 🤗<a href="https://huggingface.co/datasets/internlm/DL3DV-2k">DL3DV-2K Benchmark</a >

ETCHR-FLUX.2-klein-9B is a novel question-conditioned, reasoning-aware image editor designed to serve as a decoupled visual reasoning assistant for Multimodal Large Language Models. By decoupling the specialized image editor from the downstream understanding model, ETCHR bridges the critical bottleneck where a purely textual chain of thought fails in fine-grained focus or complex spatial transformations.

📢 News

• —🚀 [2026/05/22] We have released the training and evaluation code of ETCHR.
• —🚀 [2026/05/21] We have released the ETCHR-FLUX.2-klein-9B Model, ETCHR-SFT-400K Dataset and ETCHR GRPO-10K Dataset.

🌈 Overview

We are thrilled to introduce ETCHR (Editing To Clarify and Harness Reasoning), a novel question-conditioned, reasoning-aware image editor built on FLUX.2-klein-base-9B designed to serve as a decoupled visual reasoning assistant for Multimodal Large Language Models (MLLMs). By decoupling the specialized image editor from the downstream understanding model, ETCHR bridges the critical bottleneck where a purely textual chain of thought fails in fine-grained focus or complex spatial transformations.

</p> <p style="text-align: center;"> <img src="assets/overview.png" alt="Teaser" width="100%"> </p>

💡 Highlights

• —🔥 Decoupled & Plug-and-Play: ETCHR functions as a separate module, allowing it to assist diverse downstream MLLMs (such as Qwen3-VL-8B, Gemini-3.1-Flash-Lite, or Kimi K2.5) without requiring any task-specific fine-tuning on the understanding models themselves.
• —🔥 Naturally Reflective Pipeline: Introduces an Edit-Verify-Reason inference mechanism where the understanding model filters out noisy or flawed edits, reverting safely to the original image when verification fails.

📊 Results

We evaluate ETCHR across five distinct task families spanning fine-grained perception, chart understanding, logic reasoning, jigsaw restoration, and 3D understanding. Across all evaluated backbones, ETCHR consistently yields major improvements in Pass@1 accuracy: <p style="text-align: center;"> <img src="assets/result.png" alt="Pipeline" width="100%"> </p>

🛠️ Evaluation

Prepare your environment:

git clone https://github.com/InternLM/ETCHR.git
conda create -n ETCHR python==3.11
conda activate ETCHR
cd RL/Pref-GRPO
bash env_setup.sh fastvideo
pip install "vllm>=0.11.0"
pip install qwen-vl-utils==0.0.14

We Provide an example code running ETCHR on DL3DV-2K Benchmark in Evaluation/inference_dl3dv.py, you can start the evaluation with the following two steps:

Step 1: start a VLLM server for an understanding model (eg. Qwen3-VL-8B, Kimi K2.5, ...).

cd Evaluation
bash launch_vllm.sh

Step 2: Run ETCHR atop any understanding model

python inference_dl3dv.py

Cases

ETCHR can assist with a broad spectrum of understanding tasks, including fine-grained perception, chart reasoning, maze navigation, jigsaw puzzles, and 3D spatial understanding.

<p style="text-align: center;"> <img src="assets/case-3D.png" alt="case3D" width="100%"> </p> <p style="text-align: center;"> <img src="assets/case-jigsaw.png" alt="casejigsaw" width="100%"> </p> <p style="text-align: center;"> <img src="assets/case-maze.png" alt="casejigsaw" width="100%"> </p> <p style="text-align: center;"> <img src="assets/case-chart.png" alt="casejigsaw" width="100%"> </p>

📄 License

Our work is based on FLUX.2-klein-base-9B, so please follow FLUX Non-Commercial License.

✒️Citation

If you find this project useful, please kindly cite:

@article{zhang2026etchr,
  title={ETCHR: Editing To Clarify and Harness Reasoning},
  author={Beichen Zhang, Yuhong Liu, Jinsong Li, Yuhang Zang, Jiaqi Wang, Dahua Lin},
  journal={arXiv preprint arXiv:2605.23897},
  year={2026}
}

❤️ Acknowledgement

The base model is FLUX.2-klein-base-9B, a powerful image-to-image model.

The work is built upon <a href="https://github.com/modelscope/DiffSynth-Studio">DiffSynth-Studio</a > and <a href="https://github.com/CodeGoat24/Pref-GRPO">Pref-GRPO</a >, two excellent codebases for Diffusion models training!