📄 Paper: arxiv.org/abs/2504.02762
🌐 Project page: mdprojtex.abyildirim.com

Worked on this with Mustafa Utku Aydogdu, and had the privilege of being advised by Duygu Ceylan and Aysegul Dundar.

📌 Key contributions:
• No training or run-time optimization
• Denoising with fused latent predictions
• UV-space projection with normal-guided weighting
• Depth and lineart conditioned Stable Diffusion + ControlNet
• SOTA FID, KID, and user study results

Unlike prior methods that rely on sequential view generation or optimization, MD-ProjTex enables parallel denoising across multiple views, maintaining multi-view consistency directly in the UV texture space.

📌 Key contributions:
• No training or run-time optimization
• Denoising with fused latent predictions
• UV-space projection with normal-guided weighting
• Depth and lineart conditioned Stable Diffusion + ControlNet
• SOTA FID, KID, and user study results

Unlike prior methods that rely on sequential view generation or optimization, MD-ProjTex enables parallel denoising across multiple views, maintaining multi-view consistency directly in the UV texture space.