+
EditP23: 3D Editing via Propagation of Image Prompts to Multi-View
+
+
+ This is the official Gradio demo for EditP23, a method for fast, mask-free 3D editing that propagates 2D image edits to multi-view representations in a 3D-consistent manner. The edit is guided by an image pair, allowing users to leverage any preferred 2D editing tool, from manual painting to generative pipelines.
+
+
+
+
EditP23 requires three specific images to perform an edit. This demo automates the process, but understanding each component is key.
+
+ - Original Multi-View Image (`src_mv.png`): This is a 2x3 grid of six different views of the original, unedited object. The model uses this as the base to apply the edit consistently across all angles.
+ - Source Condition (`src.png`): This is a single, frontal view of the original object. It acts as the "before" image for the edit.
+ - Target Condition (`edited.png`): This is the "after" image. It's the same view as
src.png, but with your desired 2D modification applied. The difference between this image and src.png is what guides the 3D edit.
+
+
+
How to Prepare Your Own Images
+
You can generate the required input images using the helper scripts provided in our GitHub repository.
+
Step 1: Generate src.png and src_mv.png
+
You have two options for creating the initial views of your object.
+
+
Step 2: Create edited.png
+
Use any 2D image editor to modify your src.png. This is where your creativity comes in! For quick edits, we recommend these online tools:
+
+ - FlowEdit: Excellent for global, structural edits.
+ - Flux-Inpainting: Great for local modifications and inpainting.
+
+
+
Understanding the Parameters
+
+ n_max: Controls how many denoising steps are influenced by your edit. Higher values are needed for more significant geometric changes.tar_guidance_scale: Determines the strength of your edit. Increase this for more dramatic changes, but be aware that very high values can sometimes introduce artifacts.src_guidance_scale: Controls how strongly the model adheres to the original object's identity. This can usually be left at its default value.
+
+
Reconstructing a 3D Model
+
After this demo generates an edited multi-view image, you can use the scripts/recon.py script from our repository to convert it back into a 3D model (.obj file).
+
python scripts/recon.py path/to/instant-mesh-large.yaml --input_file "path/to/edited_mv.png" --output_dir "path/to/output/"
+