OFI-MMC

Repository Description: OFI-MMCProject Overview
This repository contains the code and data for the OFI-MMC project. We propose a Multimodal Intent Orchestrator (MIO) powered by a fine-tuned Multimodal Large Language Model (MLLM). MIO orchestrates complex multimodal intents by decomposing unstructured instructions into decoupled semantic, structural, and stylistic control streams.
Core Features
Multimodal Intent Orchestrator (MIO): Utilizing a "synthesis-distillation" strategy, the MIO precisely parses unstructured, composite user instructions into a deterministic execution blueprint. It routes decoupled control signals (semantic, structure, content, and style) into isolated channels, fundamentally preventing intent entanglement at the source.
Pairwise Stylistic Delta Coding (PSDC): We introduce the PSDC paradigm. Instead of encoding static visual appearances, PSDC uses contrastive learning in the CLIP space to capture the semantic "Stylistic Delta" between source and target domains. This mathematically purifies the stylistic representation and isolates target-independent stylistic features from redundant content information.
Orthogonal Feature Injection (OFI) Architecture: We establish a dual-pathway Orthogonal Feature Injection (OFI) architecture within the U-Net. Physical isolation is achieved as the spatial injection module anchors local geometric layouts via Cross-Attention, while the style modulation module operates on GroupNorm layers. This ensures that style-updating gradients do not project harmful interference into the structural subspace.
Performance & Advantages
Extensive evaluations on LSUN-Bedroom and CelebA-HQ datasets demonstrate that our framework effectively eliminates nonlinear cross-interference.
Our method secures a substantial 24.5% increase in the EC2EC metric (from 0.383 to 0.477), demonstrating exceptional capability in mapping intense artistic styles while maintaining pixel-level geometric fidelity.
Furthermore, our unified end-to-end architecture improves inference efficiency by nearly eight-fold compared to traditional cascaded baselines.