ComfyUI H3 Spectrum Workflow Doubled in Generation Time on RunPod Within a Week
RRSP_question1982 · reddit · 2026-09-17
A Reddit user reports their H3 + Spectrum node ComfyUI workflow on a RunPod H100 slowed dramatically within a week: a 20-step, 25s generation went from 250s to 450s; disabling Spectrum gives 330s, still above the old baseline.
Key observations:
- The slowdown is concentrated in the second half of the sampler, which used to finish in 10s.
- Suspected causes: a recent RunPod template update (config.ini much smaller), ComfyUI or Spectrum node version changes, SageAttention or RAM settings.
- LLMs were no help, citing ComfyUI changes from months ago.
More from Multimodal
- MuLaCover open-sources a controllable model that rebuilds any song into an AI cover — solyarisoftware · 2026-09-17
- Indie filmmaker spends 1.5 months crafting AI sci-fi short FIDELITY for Higgsfield festival — eraiozei · 2026-09-17
- Short film made entirely by AI agent 'astra' controlling Blender for VFX — OpenAIDevs · 2026-09-17
- Creator shares AI-generated "Christmas at the in-laws'" gag with prompts — techhalla · 2026-09-17
- Tutorial: Generate pure instrumental music with open-source YuE-2 in ComfyUI — solomars3 · 2026-09-17
- Japanese dev 3D-scans a cactus greenhouse into a 5M-splat 3DGS scene with LichtFeld — janusch_patas · 2026-09-17