Solo-built Vulkan inference engine LaForge hits ComfyUI-level SD1.5 speed: 512px in 0.8s, 402MB container
Ichidan256 · reddit · 2026-10-16
LaForge, a solo developer's Vulkan-based C++ AI inference runtime, achieved a 30-46x speedup on SD1.5 text-to-image after adding missing SPIR-V shaders and a model-residency fix on a RunPod RTX 4090: 512×512 @ 20 steps dropped from 33-37s to 0.8s, and 768×768 from 60s to 1.8-2.0s — now comparable to ComfyUI's PyTorch stack.
Other metrics: the container image is just 402 MB (vs ComfyUI slim at 7.35 GB), pod boot-to-healthy 27s, model load 10-13s on first call only. The author's goal is a lean, reusable runtime for image generation, games, and interactive AI experiences.
More from Infra
- MiniCPM-V 4.5 MoE reportedly hits 13+ tokens/s on iPhone with just 3GB memory — heyshrutimishra · 2026-10-16
- Trading 2 RTX 5090s for 4 R9700s: does doubled VRAM pay off? — poofph · 2026-10-16
- Running PolyStrata + Cline coding agent on a $0-fancy laptop at ~90 tok/s — rpwoerk · 2026-10-16
- TensorFold 1.0.12 runs 120B vision model on 1-2 DGX Sparks, adds Mac image input — HankYeomans · 2026-10-16
- vLLM brings frontier model local deployment to NVIDIA RTX Spark PCs — vllm_project · 2026-10-16
- NVIDIA signals $2B in new European AI deals at GTC Berlin, totaling ~$10B in a year — CurieuxExplorer · 2026-10-16