ComfyUI-GTE Workflow Lets You Preview Mid-Generation and Cut Exploration Costs from 25 to 2 Minutes

Think-Aioli-8203 · reddit · 2026-09-13

A developer open-sourced ComfyUI-GTE, an interactive txt2img workflow with custom ComfyUI nodes designed for exploring generation directions cheaply on rented hardware.

The core idea: place configurable checkpoints at intermediate diffusion steps. At each checkpoint you preview candidates, keep promising directions and discard the rest, then only spend remaining steps on selected trajectories. The author, who runs Flux 2 dev on a rented L40 (48GB) via RunPod, notes that generating 20 images at 20 steps normally takes 25 minutes; with this workflow good candidates surface within 2 minutes.

On top of that, the workflow can create controlled variations from a selected intermediate result, with two parameters — theta (distance from the original trajectory) and spread (diversity among sibling variations) — shaping the exploration. It also handles seed generation/visualization and plots scheduled noise remaining per step to guide checkpoint placement. Code and workflow are on GitHub.

Original post →

More from Multimodal

Multimodal channel →