TriMagic: Single-Image 3D Generation Integrating Semantic-Role Multi-Token Conditioning and Dynamic Pseudo-Multi-View Supervision

Abstract Generating 3D models from single images supports digital heritage displays, but unseen regions are unconstrained, causing Janus artifacts, cross-view conflicts, and inconsistent textures. We propose TriMagic, a single-image 3D generation framework combining semantic-role multi-tokens with dynamic pseudo-multi-view supervision. Three learnable tokens, guided toward shape, color, and texture by role-specific prompts, contrastive alignment, and residual diversity constraints, form complementary instance representations. Zero123++ provides pseudo views; against their generation errors, empirical view weights, three-stage scheduling after warm-up, and separate absolute weighting of real and pseudo views refine the generation loss. Anchored dual-view sampling pairs the real reference with one random pseudo view in each pseudo-supervision update, while a Stable Diffusion prior and multi-reference-weighted Zero123 guidance constrain novel views. On NeRF4, RealFusion15, and our Artifact, PSNR improves over Magic123 from 24.62/19.50/27.58 to 24.75/20.01/28.99 dB; Artifact LPIPS rises from 0.03 to 0.04, with gains in semantic similarity and reference-view fidelity.