Rebuilding 4 ComfyUI workflows into a zero-dependency app with DLSS 5 and VRAM streaming
MCP1964 · reddit · 2026-11-02
A long-time ComfyUI user rewrote his 4 most-used workflows from scratch into a standalone app for high-quality single renders, with no headless Comfy server or external dependencies.
Key techniques
- VRAM management: adapted comfy/modelmanagement.py partial loading to stream blocks from system RAM over PCIe 4.0, running a 22GB+ pipeline on a 16GB card and auto-unloading VRAM when idle or running local VLMs for prompt expansion.
- DLSS 5 neural rendering: integrated NVIDIA NGX DLLs via community nodes like ComfyUI-DLSS5-Enhancer for DLSS 5 NR (1x DLAA) and super resolution, replacing expensive tiled diffusion upscaling.
- Zero-graph chaining: any render can immediately start a new run with Flux Fill, Kontext, or background cutouts (adapted from RMBG/BiRefNet nodes).
Not for distribution — the value is escaping the daily maintenance of giant graphs and update breakage.
More from Infra
- OpenAI may rebrand its data centers as 'AI compute campuses' — TechCrunch AI · 2026-11-02
- $300B Louisiana project pipeline squeezes steel supply, H-beam lead times hit 9 months — MatthewChang · 2026-11-02
- Benchmark: Speedy inference engine crushes Avalon — 1.6x faster reads on Qwen3 long context — MotorNetwork380 · 2026-11-02
- vLLM serves DeepSeek-V4.1-Flash at 954K tokens/s per GPU on one GB300 NVL72 rack — bookwormengr · 2026-11-02
- Prime Mega MoE extends DeepSeek's kernel for training, up to 34% faster end-to-end MoE training — cfregly · 2026-11-02
- 2.3bpw quant update: 27B model on 12GB cards ties Qwen3.6 BF16 on LiveCode Bench at 84.3% — Brief-Tap-6616 · 2026-11-02