Official Comment · NEURIPS 2025

Thank you for your insightful questions and for ensuring the utmost rigor. We have reviewed your points carefully and offer precise clarifications below: **1) DiffusionRenderer’s Real-World Video Data vs IllumiCraft's Real-World Video Data & Novelty & Contributions** - **Thank you for pointing this out. To clarify the data settings:** DiffusionRenderer adopted the real-world videos from DL3DV-10K dataset, the DL3DV-10K dataset [1] is a curated 3D vision resource designed for scene-level learning and novel view synthesis, with multi-view coverage and sequences featuring significant camera motion and viewpoint changes to support 3D reconstruction. In contrast, IllumiPipe processes in-the-wild Pexels videos (inherently diverse, user-uploaded, real-life videos) for broad creative use. - **Regarding the novelty and contributions**, our method introduces a new exploration that unifies geometry (3D point tracks) and illumination (HDR maps) during training in a DiT-based diffusion architecture with learnable illumination tokens, an approach absent in prior work. These tokens capture fine-grained lighting information, enabling accurate relighting at inference without HDR maps while preserving temporal coherence and spatial consistency. Leveraging learned spatial and photometric priors, the model needs only an input video and text prompt at inference. To support this, our IllumiPipe pipeline automatically curated 20,170 paired in-the-wild videos from Pexels (diverse user‑uploaded videos) with synchronized relit outputs, HDR maps, and 3D tracking annotations, providing a large-scale resource for video relighting and controllable video generation. **2) Questions Regarding DiffusionRenderer Reproduction** - We adopted the academic code version from the CVPR 2025 conference (the code was released on June 11, 2025). The Cosmos version was released on June 12, 2025, these two versions are well after the NeurIPS full paper submission date (May 15, 2025). *A premier conference like NeurIPS does not require authors to compare their work with the models released after the submission date.* Our method outperforms DiffusionRenderer (CVPR 2025 academic version) in both performance and inference speed. It is worth noting that the Cosmos version are 7B models (Diffusion_Renderer_Inverse_Cosmos_7B, Diffusion_Renderer_Forward_Cosmos_7B), whereas ours is a 3B model, so the comparison is not entirely fair. **However, I report the comparison with the cosmos version in "Gentle Follow-Up on Your Latest Response \- Part 2."** - We did use the real-world LoRA (diffusion_renderer-forward-svd is chosen by default) provided with DiffusionRenderer when running the comparison. - Since the academic version of DiffusionRenderer (CVPR 2025) uses only a single HDR map for inference, we instead extract 24 HDR maps, one for each frame, from a 24-frame video. We then average the results obtained by using these 24 different HDR maps for inference, denoted as **S*** and **R***. Even