Abstract
In industries such as interior design and fashion, creating a cohesive "thematic collection"—a set of distinct objects sharing a unified aesthetic—is a fundamental yet time-consuming task. While generative AI models have shown promise in image synthesis, they often struggle with thematic consistency, frequently altering backgrounds, lighting, or specific material textures when generating new objects. This paper introduces iCONTRA (interactive CONcept TRAnsfer), a zero-shot system designed to transfer aesthetic concepts from a single reference image to new target objects while preserving the original scene's context. By leveraging a novel "fade-in" attention mechanism and a grid-based user interface, iCONTRA enables designers to iteratively refine collections without the need for model fine-tuning.
Introduction
Professional designers often find inspiration in a single "anchor" object—a vintage leather sofa, for instance—and must then design a suite of matching items like beds, tables, or wardrobes. Achieving this level of consistency using standard generative tools like DALL-E or Midjourney is notoriously difficult. These tools often suffer from "concept drift" or fail to maintain the environmental context (lighting, floor textures, and shadows) of the original inspiration. Furthermore, existing solutions for style transfer often require computationally expensive fine-tuning on specific datasets, which is impractical for rapid design workflows.
iCONTRA addresses these challenges by providing a zero-shot framework that allows for instant concept transfer. It ensures that the generated objects inherit the material and stylistic properties of the source while remaining perfectly integrated into the original background.
System Design and User Interface
Unlike many AI tools that rely on a conversational chatbot interface, iCONTRA utilizes a specialized web-based platform built with Gradio. The interface is centered around a grid of six "generation cells." This design choice is intentional: it allows users to keep the reference image as a constant visual anchor while exploring multiple variations in parallel. By avoiding the linear nature of chat-based generation, the system prevents the AI from "forgetting" the original reference as the session progresses, a common issue in iterative prompting.
Technical Methodology
The technical core of iCONTRA is a zero-shot image editing algorithm built upon Stable Diffusion 1.5. The process involves three critical stages:
- Deterministic Inversion: The system uses DDIM inversion with null-text guidance to reverse-engineer the original reference image into the AI's latent space. This ensures that the starting point for generation is an exact mathematical representation of the source image.
- Automated Masking: To preserve the environment, the system employs libraries such as
RembgorLangSamto automatically identify and mask the foreground object. This allows the AI to replace only the specific object of interest while keeping the background, shadows, and lighting completely intact. - FiOII Attention Mechanism: The "Fade-in Object-to-Image Interpolation" (FiOII) is the system's primary innovation. During the denoising process, the system gradually introduces features from the source object into the new target. By interpolating intermediate features, the AI can query the specific textures (e.g., the grain of wood or the patina of leather) of the source and apply them to the structure of the target object. This results in a more natural "inheritance" of style rather than a simple overlay.
Evaluation and Results
The researchers conducted a pilot study involving design students to evaluate the system's efficacy across various tasks, such as generating a matching ring from a watch or a bed from a sofa.
Key Findings:
- High User Satisfaction: Participants reported a satisfaction level exceeding 7 out of 10 for the generated images, noting that the results felt like they belonged in the same professional collection.
- Contextual Integrity: Compared to baselines like BingAI (DALL-E 3), iCONTRA demonstrated superior performance in maintaining the "scene." While BingAI often generated unrealistic environments or altered the background entirely, iCONTRA kept the room and lighting consistent.
- Efficiency: Users found they could achieve their desired results with fewer prompts because the system's architecture naturally prioritized the stylistic features of the reference image.
Discussion and Limitations
While iCONTRA represents a significant step forward in thematic design, several limitations remain. The system currently struggles with regional editing, such as changing a small detail like a zipper on a bag, because the attention mechanism tends to focus on the most salient (prominent) parts of the object. Additionally, the system is sensitive to the initial layout; it is difficult to generate the same object from a radically different viewpoint or in a significantly different pose than what the initial prompt suggests.
Future work will focus on integrating user-defined masking for more granular control and developing features to export these AI-generated concepts into professional design software like Adobe Illustrator or Photoshop.
Conclusion
iCONTRA demonstrates that by combining clever attention mechanisms with a designer-centric interface, generative AI can become a precise tool for thematic consistency. By eliminating the need for fine-tuning and prioritizing context preservation, iCONTRA offers a viable path toward AI-assisted professional design workflows.
Summary updated