🗺️ Gimbal Manifold Explorer

Instead of generating nine images with nine random seeds and hoping the dice scatter somewhere useful, Gimbal Manifold Explorer generates nine arranged images: a 2D grid that spans two independent directions through latent space at once. One axis might be "warm earthy → cool obsidian," the other "close → distant," and the node hands you every combination as a single batch. You get a map of your idea's neighborhood instead of a grab bag.

It's the survey instrument of the Gimbal-comfy suite by Form & Noise, and the workflow pattern is the pack's signature: let the Cross-Modal Bridge produce the two axis vectors from text, hand them here as x_vector and y_vector, and read the whole terrain off one output.

How it works

You give it a center_latent (your starting point), an x_vector and y_vector (the two directions to explore), and grid dimensions. The node builds offsets across each axis - grid_size_x positions from one side of the x-vector to the other - and for every (x, y) cell it displaces the center latent by x_strength × offset along x and y_strength × offset along y. The interpolation_mode picks whether that displacement is a plain Linear nudge or a Slerp walk along the geodesic arc (Slerp_Origin for the simpler anchor). normalize_vectors (on by default) keeps the axis vectors unit-length so strengths behave consistently, and the clamp toggles keep wild outputs from flying out of range.

The result comes out as latent_batch - a single LATENT whose batch dimension is grid_size_x × grid_size_y, ready for one KSampler pass. You also get grid_meta (a DICT) and grid_report (a STRING describing the layout), and the optional mu_override lets you supply a custom centroid anchor for μ-centered Slerp instead of defaulting to the center latent.

The inputs that matter

• grid_size_x / grid_size_y - 1 to 16 each. Keep them small at first; 3×3 is the tutorial default and a fine place to start.
• x_strength / y_strength - how far each axis reaches. −10 to 10, and negative strengths explore the opposite direction, which is genuinely useful for "warm → cool" sliders.
• interpolation_mode - start with Linear; reach for Slerp when midpoints look washed out.

Then decode the batch through a KSampler → VAEDecode → Gimbal Grid Stitch to see the whole map as one contact sheet.

Common issues

Mind VRAM: a 16×16 grid is 256 latents through the sampler at once - that's a lot even on a big card, so chunk your grid or refine at lower resolution. And the pack's golden rules still apply: keep the refinement pass at denoise ~0.5 with modest CFG, or the sampler overwrites the carefully plotted grid.

Install

# ComfyUI Manager → search "Gimbal"
# or:
cd ComfyUI/custom_nodes
git clone https://github.com/FormAndNoise/Gimbal-comfy
pip install -r Gimbal-comfy/requirements.txt

Restart ComfyUI; it's under Add Node → Gimbal/Flight Instruments. Dependencies are just torch, numpy, pillow.