Execution Cost Estimator (Arctenox's Essentials)

The Execution Cost Estimator is the "will this OOM?" node. Before you run a generation, it estimates the VRAM you'll need, roughly how long the run will take, and how efficiently your batch size uses the GPU - then prints a human-readable report. It's a planning tool, not a profiler, and the author is upfront that the numbers are guidance, not guarantees.

It's an output node (cost_report, a STRING), meaning you place it as a terminal in the graph and read its result after execution. It's a natural companion to the pack's VAE Encode + Dimensions, which can hand it the real resolution instead of you typing it.

Inputs you'll actually touch:

• width, height, batch_size, steps - your generation's geometry. Defaults are 960×1280, batch 1, 25 steps.
• model_type - the architecture, as a dropdown: SD 1.5 (4GB), SD 2.1 (6GB), SDXL (8GB), Flux.1 Dev (24GB), Flux.1 Schnell (12GB), or Custom Model with a custom_model_size_gb field. Pick honestly; the estimate is only as good as this.
• precision - fp32 / fp16 / bf16 / fp8. This is where the estimate actually earns its keep, because memory scales hard with precision. At SDXL sizes, fp16 vs fp8 can be the difference between fitting and not.
• enable_vae - whether to include VAE decode in the estimate. If your graph saves the latent and decodes later, flip it off.
• device - auto, cuda, mps, or cpu. Let it stay on auto unless you're deliberately planning for a different target.

There are also optional model and vae inputs. Wire the actual MODEL in and the estimator can read real architecture and size details instead of trusting your dropdown selection - worth doing for one-off serious checks.

How the math works matters less than what it's for. It estimates: model resident size at the chosen precision, latent tensor size for your resolution and batch, and a per-step time figure, then combines them into VRAM pressure warnings, a total runtime estimate, and a batch-efficiency assessment (bigger batch usually amortizes fixed overhead better - up until you run out of memory, which is precisely what this is meant to warn you about before it happens).

The honest framing: "estimate" is doing real work in that name. It won't know your exact GPU, your CFG's sampling cost, or your scheduler's step count quirks. Use it as a pre-flight check when you're scaling up a resolution or batch size you've never tried - the alternative is learning about OOMs mid-queue. For day-to-day generation where you already know your card's limits, it's mostly a curiosity.

Install. Ships in Arctenox's Essentials - ComfyUI Manager → search "Arctenox's Essentials", or:

cd ComfyUI/custom_nodes/
git clone https://github.com/Arctenox/Arctenoxs-Essentials_ComfyUI

Restart ComfyUI. Dependencies are just torch, numpy, and optional psutil (which this node can use for live memory monitoring). No downloads. The README marks the pack deprecated while the author remasters it - fine for a utility you can eyeball in a minute, but don't bet a critical workflow's stability on a pack that's being rebuilt.