ComfyUI queue stuck? A maintainer's checklist to separate validation, node failures and lost progress
fluxdraw · reddit · 2026-10-06
A ComfyUI maintainer shares a practical troubleshooting checklist built around one principle: identify which stage failed before reinstalling anything.
- Preserve evidence first: export the workflow JSON and copy the full traceback with node IDs; screenshots often miss the relevant exception
- Check validation before blaming CUDA: missing model filenames, empty required inputs, missing node classes and incompatible sockets can stop a graph pre-execution; API clients should inspect nodeerrors in the response body
- Separate execution from display: prompts go over HTTP, progress over WebSocket — a frozen progress bar may mean the render is still running; check backend logs and output folders, and ensure reverse proxies support WebSocket
- Isolate custom-node conflicts: test with core nodes only, then restore third-party nodes in small groups; install deps in the Python env ComfyUI actually uses
- Treat VRAM errors specifically: note whether allocation failed during model load, sampling or VAE decode; try batch size one with unchanged model/seed
- Write reproducible bug reports: install method, OS/GPU/driver, version numbers, minimal workflow
A longer symptom-by-symptom guide is linked in the post.
More from Multimodal
- Apple researcher lands two NeurIPS 2026 papers on normalizing flows, including Normalizing Trajectory Models — thoma_gu · 2026-10-06
- SemanTok: 49M autoregressive video model beats SOTA rivals 47x its size — CSProfKGD · 2026-10-06
- Magnific teases October 8 launch, simple prompts already yield impressive results — aziz4ai · 2026-10-06
- Gaussian GRPO normalizes multimodal RL reward distributions, boosting OpenVLThinker v2 — kaiwei_chang · 2026-10-06
- One Prompt, $1 Budget: Agent Produces a 30s Video via Lora Pilot MCP — no3us · 2026-10-06
- Dev uses GitHub Copilot app to generate a rebuildable 60s hype video via PR — DanWahlin · 2026-10-06