Open lalalune opened 1 year ago
Thanks for the encouragement! The issue of front face hallucination (Janus) remains unresolved in many cases, and we are exploring some solutions less reliant on the language model than view-dependent prompting.
SJC does not pursue surface modeling the same way that DreamFusion does (we hope to gradually develop the skills). The latitude a full volumetric representation has makes the burger meat looks more realistic (even without view dependencies in a nerf such as ours). To be fair, this is more of a trade-off.
We will continue to try to improve the quality & robustness of the codebase.
Just wanted to say--
This is incredibly interesting work and seems to have a robustness to Janus issues (front identity applied to many angles) and quality of color that we've not yet seen from open source implementations of Dreamfusion.