LTFloat_Step_0001

LTFloat_Step_0001 is the step helper with a 0.001 increment, and it's the one the Latent Tools README actually demonstrates: generate images with a fixed seed Gaussian noise, starting with σ between 0.8 and 1.2, stepping by 0.001 per run. That's the pack's signature move - 400 images, a gentle gradient across the standard deviation, and the output becomes an animation of how noise shape changes an image.

Of the thirteen step siblings (0.0001 through 1.0), this is the one most people land on first. 0.001 per step is fine enough that consecutive images look related, but coarse enough that a reasonable batch actually gets somewhere. It's the sweet spot for sweeping a noise parameter like σ or μ.

How it works

Keep the honest model in your head: the node does not auto-increment. It takes the value you type, rounds it to five decimals, and returns it as a FLOAT and a STRING. The step is the widget's spinner increment and the label for the constant. Between runs, you bump the value by 0.001 (or use the spinner's nudge), change nothing else, and let the fixed seed do the rest.

The inputs and outputs

• value - FLOAT, default 0, range ±1,000,000, step 0.001.
• FLOAT - the value, rounded to five decimals.
• STRING - the value as text, ready for a filename prefix.

Wire the FLOAT into the mean or std input of LTGaussianLatent and the STRING into a Save Image filename, and every image out of the sweep is labeled with its own parameter. That's the whole workflow: constant → noise → sampler → labeled save.

Where it shines (and doesn't)

This is the step to reach for when the README's own 0.8–1.2 σ sweep is the shape of what you want. If you need even finer per-frame drift for video, drop to 0.0001; if 0.001 feels like watching paint dry, the 0.01 sibling cuts the batch size by ten. Start here - it's the calibrated default.

Installing it

Part of xl0's Latent Tools pack. ComfyUI Manager → search Latent Tools → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/xl0/latent-tools

Restart, and it's under LatentTools. The pack's only dependency is lovely-tensors (Manager installs it; manual cloners may need pip install lovely-tensors). No model downloads.