💻 computer science

CatalogStitch: Dimension-Aware and Occlusion-Preserving Object Compositing for Catalog Image Generation

CatalogStitch introduces a set of model-agnostic techniques that automate dimension-aware mask computation and occlusion-preserving restoration, transforming generative object compositing into a practical, user-friendly tool for production catalog image generation by eliminating tedious manual adjustments.

Original authors: Sanyam Jain, Pragya Kandari, Manit Singhal, He Zhang, Soo Ye Kim

Published 2026-04-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sanyam Jain, Pragya Kandari, Manit Singhal, He Zhang, Soo Ye Kim

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a digital interior designer trying to swap out a piece of furniture in a photo. Maybe you want to replace a wide, low coffee table with a tall, skinny floor lamp. Or perhaps you want to put a new sofa in a room, but there's a potted plant sitting right in front of it.

In the world of Generative AI (the technology that creates images), doing this swap used to be a nightmare for non-experts. The AI would try to force the new item into the old space, often squishing the lamp flat like a pancake or erasing the plant entirely.

Enter CatalogStitch. Think of it as a "smart assistant" that sits between you and the AI, fixing the mistakes before they happen so you don't have to.

Here is how it works, broken down into simple concepts:

1. The "Stretchy Pants" Problem (Dimension-Aware Masking)

The Old Way: Imagine you have a pair of jeans (the old product) and you want to replace them with a long, flowing dress (the new product). If you just tell the AI, "Put the dress in the spot where the jeans were," the AI might try to stretch the dress to fit the exact shape of the jeans. The result? A distorted, weird-looking dress that looks like it's been pulled by a giant.

The CatalogStitch Solution:
CatalogStitch acts like a tailor. Before the AI even starts working, it measures the new item.

If the new item is taller or wider than the old one, the tailor automatically expands the "cutting zone" (the mask) to fit the new shape.
It keeps the item centered in the room but gives it the breathing room it needs to look natural.
The Result: The lamp stands tall and straight, not squashed. You don't need to manually adjust anything; the system just knows to make the space bigger.

2. The "Magic Eraser" Problem (Occlusion Preservation)

The Old Way: Imagine you want to replace a sofa, but there is a coffee table and a vase sitting in front of it. When the AI tries to swap the sofa, it often gets confused. It might try to "paint over" the vase to make room for the new sofa, or it might hallucinate a weird, blurry version of the vase. It's like trying to change the background of a photo while accidentally deleting the foreground.

The CatalogStitch Solution:
CatalogStitch uses a three-step "Save and Restore" trick:

The Snapshot: Before the AI touches anything, it takes a high-definition "snapshot" of the things blocking the view (the vase, the table). It saves these pixels in a safe locker.
The Clean Sweep: It tells the AI, "Okay, pretend those objects aren't there." It digitally erases them to create a clean, empty background.
The Swap & Restore: The AI swaps in the new sofa on the clean background. Once that's done, CatalogStitch takes the "snapshot" from the safe locker and pastes the vase and table back exactly where they were, pixel-perfect.

The Result: The new sofa looks perfect, and the vase in front of it looks exactly like it did in the original photo—no blurring, no weird artifacts.

Why This Matters for Everyone

Previously, using AI to create catalog images (like for Amazon or IKEA) required a team of experts to manually fix every mistake. It was slow and expensive.

CatalogStitch changes the game by making the process model-agnostic. This means it doesn't matter which specific AI engine you use (like ObjectStitch, OmniPaint, or InsertAnything); CatalogStitch wraps around them like a protective shell, making them all smarter and more reliable.

The Big Picture:

Before: You are a mechanic trying to fix a car engine with your bare hands, struggling with every bolt.
After: You are a driver. You just say, "I want to go to the beach," and the car (CatalogStitch) automatically adjusts the suspension, changes the tires, and navigates the road for you.

By automating the boring, technical fixes, CatalogStitch lets marketers and designers focus on the creative part: choosing the right products and the right backgrounds, without needing to be coding wizards or image-editing experts. It turns a complex, error-prone process into a simple, one-click workflow.

1. Problem Statement

Generative object compositing models (e.g., ObjectStitch, OmniPaint, InsertAnything) have advanced the ability to insert objects into scenes seamlessly. However, they fail in real-world e-commerce catalog workflows due to two critical gaps between research assumptions and production needs:

Challenge 1: Product Dimension Mismatch: Existing methods often force a new product into a target mask defined by the original object. If the new product has a different aspect ratio (e.g., replacing a wide sofa with a tall lamp), the result is severe geometric distortion (stretching or squashing) to fit the original mask.
Challenge 2: Occlusion Destruction: Real catalog images often feature foreground elements (plants, vases, furniture) partially occluding the target product. Standard compositing replaces the entire target region, destroying or hallucinating these foreground elements, requiring tedious manual post-editing to restore them.

Current workflows require significant manual intervention (mask adjustment, scaling, and pixel-level restoration), making them impractical for non-experts and slow for high-volume production.

2. Methodology: CatalogStitch

The authors propose CatalogStitch, a set of model-agnostic techniques that wrap around any existing generative compositing model. The system operates as a preprocessing and postprocessing pipeline, requiring no model fine-tuning.

A. Dimension-Aware Mask Computation (Preprocessing)

This module automatically adapts the target mask to the aspect ratio of the new product while maintaining the original spatial location.

Logic: It calculates the aspect ratio ($AR$) of the original target mask ( $AR_t$ ) and the new product ( $AR_p$ ).
Thresholding: If $|AR_t - AR_p| > \tau$ (where $\tau=0.06$ ), the system triggers an expansion.
Algorithm:
1. Compute the centroid of the original mask to preserve positioning.
2. Calculate optimal new dimensions ( $w^*, h^*$ ) based on the product's aspect ratio, anchoring either on width or height depending on which dimension is smaller to minimize scene intrusion.
3. Generate a new rectangular mask centered on the original centroid, clipped to image boundaries.
Result: The generative model receives a mask that naturally accommodates the product's true proportions, preventing distortion.

B. Occlusion-Aware Hybrid Restoration (Postprocessing)

This module ensures that foreground elements occluding the target product are preserved with pixel-perfect fidelity using a deterministic "cache-and-restore" approach rather than generative prediction.

Five-Stage Pipeline:
1. Segmentation: Identify foreground instances overlapping the target region using entity segmentation (e.g., EntitySeg).
2. Caching: Extract and cache the exact RGB pixels and alpha masks of these occluders from the original background image.
3. Inpainting: Apply generative inpainting to the background image to remove the occluders, creating a clean, unoccluded background.
4. Compositing: Run the generative compositing model on the clean background using the dimension-aware mask. This produces a seamless integration without occlusion artifacts.
5. Recomposition: Alpha-composite the cached original occluder pixels back onto the generated result.
Result: The final image contains the newly generated product integrated into the scene, with the original foreground elements restored exactly as they appeared, preserving lighting, texture, and geometry.

3. Key Contributions

Dimension-Aware Mask Algorithm: An automated geometric solution that eliminates manual mask resizing and prevents aspect-ratio distortion.
Occlusion-Aware Hybrid Restoration: A novel hybrid approach combining generative AI (for background cleaning and harmonization) with deterministic operations (pixel caching) to guarantee the preservation of occluding elements.
CatalogStitch-Eval Benchmark: A new evaluation dataset containing 58 examples (35 dimension-mismatch, 23 occlusion-heavy) with ground truth masks, metadata, and HTML viewers for rapid inspection.
Cross-Model Validation: Empirical proof that these techniques improve performance across three state-of-the-art models: ObjectStitch, OmniPaint, and InsertAnything.

4. Experimental Results

The authors evaluated their method against baselines using five metrics: Aspect Ratio (AR) Error, FID (Fréchet Inception Distance), CLIP-score, DINO-score, and Occluder PSNR.

Dimension Preservation: AR Error dropped significantly from ~30% (baselines) to ~4–5% across all models.
Image Quality: FID scores improved substantially (e.g., InsertAnything improved from 105.99 to 77.72), indicating more realistic composites.
Semantic & Structural Fidelity: CLIP and DINO scores increased, showing better alignment with the reference product and structural integrity.
Occlusion Restoration: For occlusion scenarios, the method achieved high Occluder PSNR (up to 27.54 for InsertAnything + Ours), proving the effectiveness of the pixel-restoration technique compared to the low fidelity of baselines (e.g., 11.60 for ObjectStitch baseline).
Consistency: The improvements were consistent regardless of the underlying generative model, confirming the techniques address systematic failure modes rather than model-specific quirks.

5. Significance and Impact

Production Readiness: CatalogStitch bridges the gap between experimental AI research and industrial e-commerce needs. It transforms compositing from a manual, expert-heavy task into an automated, user-friendly workflow.
Human-AI Interaction Shift: It shifts the user's role from low-level pixel manipulation (masking, scaling) to high-level creative decisions (selecting products and backgrounds).
Hybrid Design Philosophy: The paper demonstrates that combining generative flexibility (for shadows, lighting, and harmonization) with deterministic reliability (for exact pixel preservation and geometric constraints) is the optimal path for professional-grade image generation.
Accessibility: By removing the technical overhead, it democratizes high-quality catalog creation for marketers and designers without specialized AI skills.

In conclusion, CatalogStitch provides a robust, plug-and-play framework that solves the two most critical barriers to deploying generative compositing in real-world retail environments: geometric distortion and occlusion loss.