Latent Transform Sampler 🔄 (Legacy)

It's not a different node

NTransformSampler - shown in ComfyUI as "Latent Transform Sampler 🔄 (Legacy)" - is not a separate implementation. Open the pack's source and the registration table maps both class names to the very same class:

NODE_CLASS_MAPPINGS = {
    "LatentTransformSampler": LatentTransformSampler,
    "NTransformSampler": LatentTransformSampler,  # Backward compatibility
}

The "(Legacy)" label is the giveaway. This alias exists so workflows saved under the old class name keep loading after the rename, and it's the only thing the legacy version adds. Identical inputs, identical behavior, identical output. If you see it in an old workflow, you can leave it alone or swap in the current node - you get the same image either way.

What it does, briefly

It's a KSampler replacement that applies a geometric transform - shift, mirror, or rotation - to the latent a few times mid-denoise and lets the model redraw around the change. The full explanation lives on the LatentTransformSampler page, and everything there applies here unchanged, because this node is literally the same code path. transform_count, transform_type, distribution, shift_pixels_x/shift_pixels_y, manual_steps and transform_sequence all behave exactly as described there, and it outputs a single samples (LATENT) for VAE Decode. Yes, debug_mode defaults to on here too, so expect the schedule printed to your console on first run.

Install

One pack installs both names:

cd ComfyUI/custom_nodes
git clone https://github.com/pizurny/Comfyui-Latent-Transform-Sampler

Restart ComfyUI and both nodes appear under sampling/transform. No models to download and no extra dependencies beyond torch and numpy, which ComfyUI already ships. ComfyUI Manager finds the pack as "Latent Transform Sampler" if you prefer that route.

Which one should you use?

The current one. The legacy name exists purely for backward compatibility, and new workflows are better off using LatentTransformSampler so you're not dragging an alias around. If an old workflow you love still references NTransformSampler, there's zero urgency to change it - but the next time you rebuild or share that workflow, the current node is the clean choice. The gotchas are shared too: manual step lists get truncated by transform_count, random modes ignore your seed, and shifts wrap across the latent's edges.