GIDE

Diffusion Large Language Models (DLLMs) show strong multi-modal generation ability, but precise training-free image editing remains difficult due to their discrete token space.
To address this, we propose GIDE (Grounded Inversion for DLLM Image Editing), a unified framework with three stages: grounding, inversion, and refinement.
GIDE introduces a discrete noise inversion mechanism to preserve structure and background consistency while supporting diverse editing instructions (text, point, and box).
We also introduce GIDE-Bench, a benchmark with 805 compositional editing scenarios. Experiments show clear improvements over previous training-free methods, including strong gains in Semantic Correctness (SC) and Perceptual Quality (PQ).

The following table presents quantitative comparisons on GIDE-Bench, where GIDE significantly outperforms state-of-the-art training-free methods, improving Semantic Correctness by 51.83% and Perceptual Quality by 50.39%.

conda create -n GIDE python=3.12
conda activate GIDE

git clone https://github.com/Zivenzhu/GIDE.git
cd GIDE
pip install -r requirements.txt

You are also welcome to access our GIDE-Bench on Hugging Face:

You need to install both SAM 2 and SAM 3.

git clone https://github.com/facebookresearch/sam2.git
mv sam2 sam2_repo
cd sam2_repo
pip install -e .

git clone https://github.com/facebookresearch/sam3.git
mv sam3 sam3_repo
cd sam3_repo
pip install -e .
pip install -e ".[notebooks]"

cd evaluation
python3 evaluate_ours.py --lambda1-ratio 0.2 --use-grounding-module

python3 evaluate_ours.py --lambda1-ratio 0.2 --use-grounding-module --w-o-inversion

python3 evaluate_ours.py --lambda1-ratio 0.2

python3 evaluate_ours.py --lambda1-ratio 0.2 --use-grounding-module --w-o-refinement-segement

python3 evaluate_ours.py --lambda1-ratio 0.2 --use-grounding-module --w-o-intrinsic-refinement

python3 evaluate_ours.py --lambda1-ratio 0.2 --use-grounding-module --w-o-residual-recovery

You can adjust the mixing coefficient $\lambda$, which controls the strength of the introduced inversion residual $\boldsymbol{z}_t$.

For example, set $\lambda$ to 0.4:

python3 evaluate_ours.py --lambda1-ratio 0.4 --use-grounding-module

For example, to evaluate GPT-Image-1:

python3 evaluate_gpt.py

Other methods can be evaluated in a similar way.

After obtaining model responses, run:
This step computes the Semantic Correctness (SC) and Perceptual Quality (PQ) scores for each image.

python3 metrics.py

Then compute the average score:

python3 calculate_score.py