Mastering ComfyUI — Part 9: Performance & Troubleshooting

Previous: Part 8 — Workflow Management

Where VRAM actually goes

Every piece of a workflow that's loaded onto the GPU competes for the same pool of VRAM:

• The checkpoint itself (MODEL + CLIP + VAE) — several GB depending on the model; Flux and other larger architectures cost more than SDXL-family checkpoints.
• Any LoRAs applied on top (small individually, but they still occupy memory once loaded).
• Any ControlNet models in use.
• Any upscale models in use.
• The actual working memory needed during sampling, which scales with resolution and batch size — this is the part that changes run-to-run even with the same models loaded.

When you run out, ComfyUI throws an out-of-memory error rather than silently degrading — so the goal is managing the inputs you control (resolution, batch size, how many models are loaded at once) rather than reacting after a crash.

The levers you actually control

Batch size — the most direct trade: batch_size on Empty Latent Image (or wherever your batch is set) multiplies VRAM cost roughly linearly. Generating 4 images at once costs roughly 4x the working memory of generating 1. If you're pushing resolution or a heavy model and hit memory limits, batch size is the first thing to drop back to 1.

Resolution — also scales memory cost, though less linearly than batch size. This is another reason the "generate at native resolution, then upscale" pattern from Part 6 matters for more than just quality — a native-resolution generation followed by an upscale pass is often cheaper in peak memory than one enormous single-pass generation, in addition to producing better output.

Tiled VAE decode — VAE Decode (and VAE Encode) can, on some ComfyUI builds/nodes, run in tiled mode (VAE Decode (Tiled)), processing the image in smaller chunks rather than all at once. This trades a small amount of speed for a real reduction in peak VRAM specifically during the decode step — worth reaching for if you're comfortably fitting a generation in memory but hitting an out-of-memory error specifically at the decode stage, which is a distinct symptom from running out of memory during sampling itself.

How many models are loaded simultaneously — a workflow with a checkpoint, two LoRAs, a ControlNet model, and an upscale model all active at once has all of them resident in VRAM together. If you're memory-constrained, look for the actual heaviest workflow you're running and count what's loaded, rather than assuming the checkpoint alone is the whole cost.

Monitoring GPU usage

Keep an eye on actual usage rather than guessing:

• nvidia-smi (or nvidia-smi -l 1 for a live-updating view) from a terminal shows current VRAM usage and GPU utilization directly — the most reliable source of truth for what's actually happening on the card right now.
• ComfyUI's own Console panel (Part 1) will often show a specific out-of-memory traceback naming which allocation failed, which is more informative for diagnosing where in the pipeline memory ran out than nvidia-smi alone.

Consolidated error reference

A single place to check first when something goes wrong, gathering everything covered across this series:

Symptom	Likely cause	Where it's covered
"Missing Models" / red-outlined node	Filename mismatch between workflow and what's actually in your models folder	Part 3
A specific input dot shows unconnected, run refuses	A required input was never wired	Part 3
Inpainted region doesn't change	Empty mask, or wrong source image reselected after painting	Part 5
Positive prompt seems ignored	Stale wiring still pointing at an old/unused node chain	Part 5
LoRA "does nothing"	Missing trigger word, or strength set too low	Part 4
Output distorted/oversaturated after adding a LoRA	Strength set too high	Part 4
Two similar subjects merge into one hybrid figure	Fundamental limit of regional conditioning; try distinct poses or inpainting instead	Part 5
Hires-fix pass makes the image worse	Second-pass denoise set too high	Part 6
Upscale node errors on mismatched size/channels	Feeding it a latent instead of a decoded image, or vice versa	Part 6
ComfyUI fails to start after installing an extension	A custom node pack failed to import — check console traceback for which one	Part 7
"Missing node type" loading someone else's workflow	A custom node pack referenced isn't installed — use Manager's resolver	Part 7
Out-of-memory error	Batch size, resolution, or total loaded models exceeding available VRAM	This part
Out-of-memory specifically at decode	Try tiled VAE decode	This part

Next: Part 10 — Looking Ahead.

Previous: Part 8 — Workflow Management