Thank you to everyone providing logs and feedback in the discussions and via our support email after the recent release of V2.0. To all the early adopters, solo-devs, and indie studios who jumped into the trenches with us—thank you. We have been reading every log, debugging every crash, and tearing down the engine to build something better.
Jupetar V2.5 is officially live, and it addresses the biggest pain point you’ve reported: Pipeline Stability.
We have entirely rewritten the C# to Python inter-process bridge to eradicate the agonizing file-lock timeouts and "black box" crashes.
If you recently upgraded to an RTX 50-series graphics card, you likely ran into a CUDA error: no kernel image is available crash during the geometry phase. We have completely recompiled the underlying PyTorch, xFormers, and NVIDIA Nvdiffrast C++ backend directly against the latest CUDA 12.4 toolkits. The engine now natively recognizes the new sm_100 architecture, and 50-series cards will now tear through generations flawlessly.
Fixed a silent dependency break that was bypassing our neural upscaler. Your texture atlases are now properly passing through Real-ESRGAN to physically hallucinate high-resolution fabric weaves, pores, and metal grain.
Implemented a brand new mathematical curvature engine. Instead of just painting dark colors over the model, Jupetar now intelligently calculates the physical geometry of your generated mesh. It dynamically lowers the roughness on sharp outer edges (creating realistic shiny metal wear) and injects realistic Ambient Occlusion (AO) shadows deep into crevices.
The RMA (Roughness/Metallic/AO) packer has been upgraded to properly bind these new procedural metallic channels to your export, meaning your .glb files will react beautifully to dynamic lighting in Unreal, Unity, or Blender.
Fixed a deep C++ PyTorch wrapper issue that was causing a RuntimeError: Tried to instantiate class 'imageSuperNet' crash for some users right as the texturing phase began.
The Interactive Clay Viewport: No more waiting 10 minutes blindly! The exact microsecond Jupetar finishes the Stage 1 geometry sculpt, the C# engine now intercepts the raw .glb and pushes a live, interactive 3D blockout directly to your viewport. You can inspect the topology and wireframe while the heavy 4K texturing runs silently in the background.
The "Omni-Bake" Rasterizer Fix: We patched a critical math collision in the C++ nvdiffrast engine that was causing the pipeline to fatally crash during the Multi-View texture bake. Texturing is now fully stabilized for 10GB+ VRAM cards.
The "Anti-Obsidian" Material Lock: We caught a bug where highly reflective procedural materials were causing the viewport to render models as pitch-black "wet obsidian." V2.5 introduces a dynamic material lock in the UI, complete with active PBR sliders (Roughness/Metalness) so you can inspect your models accurately.
Ghost-Texture Eradication: Fixed a persistent caching bug where previous generations' textures would randomly bleed onto new "Geometry-Only" exports.
As always, we thank the community for their patience while we fortified the foundation. We are just getting started.
— Odyssey Game Studios