Current local AI capabilities on this machine: - LLM inference (KoboldCpp / Vulkan / Silly Tavern) - Speech synthesis (AllTalk / ROCm) - Face swapping/age mod/enhance etc (FaceFusion / ROCm) - Image-to-3D reconstruction (2nd ComfyUI / TRELLIS.2 / ROCm) - Image-to-video generation (LTXv2 atm / ComfyUI / ROCm) - Text-to-image generation (SDXL / ComfyU / ROCm)
Update today: text to picture
edit: Trellis2 / CuMesh ROCm – Update - Finally managed to render a +15000 tokens low poly model (<10000 polys) with remeshing (high polycounts already worked before).
Latest results: - 16,698 tokens processed - Shape SLat HR: 8m 35s - Texture SLat: 4m 41s - Total generation time: 18m 59s - Initial geometry: 9.31M vertices / 22.9M triangles - Final mesh: ~9,515 triangles - Result: Highly detailed, textured low-poly model with very few holes! - GPU: AMD RX 6950 XT / ROCm 6.4.4
Latest patches: - keep_models_loaded = false (did not help with remesh oom) - Enabled CuMesh's internal timing diagnostics (simplify_step(..., True)), including GPU synchronization. - Added automatic intermediate mesh capture before simplification for standalone debugging. - No modifications to the actual CuMesh simplification algorithm.
Acknex umgibt uns...zwischen Dir, mir, dem Stein dort... "Hey Griswold ... where u gonna put a tree that big ?" 1998 i married my loved wife ... Sheeva from Mortal Kombat, not Evil-Lyn as might have been expected rayp.flags |= UNTOUCHABLE;