Latent interpolate transform (one time)

OneTimeLatentInterpolateTransform is the one-shot version of Latent interpolate transform: instead of blending the denoised prediction toward a reference latent across a window, it does one weighted blend at exactly one step. It's the subtle member of the one-time family - the effect is mild by design, which makes it both easy to ignore and easy to stack.

The inputs

• latent - the reference latent to blend toward (upscaled to match if needed).
• step - the step to fire on (1-based).
• factor - the blend weight, 0 to 1. 0 does nothing, 1 replaces the prediction with the reference, 0.5 is halfway. Default 0.5.
• multiplier - scales the blended result.

Output is a TRANSFORM for a TSampler, TSampler Advanced, or Transform Hijack-patched sampler.

Why one blend step is worth it

Because the sampler propagates the blend. A single interpolation at step 6 of 20 tilts the trajectory toward the reference's character, and every subsequent step builds on that tilt - so a small, one-step nudge ends up influencing the whole rest of the run. That's the trick: you don't need to blend continuously to steer the result. One early blend at factor 0.4 can shift a generation's mood noticeably without the "reference takeover" that a continuous blend tends to produce.

It's also the cleanest node in the pack for learning what interpolation does. The windowed version asks you to think about windows and offsets; this one asks for one number and one step. Blend once, compare to a no-blend run of the same seed, and you'll immediately see how much a single tilt matters.

Practical notes

• Low factor (0.2–0.4) at an early step = gentle mood steering; high factor late = a strong final-character stamp.
• A one-step blend is mild enough that you can stack several at different steps (via Combine transforms) to fake a ramp: early low-factor tilt, later higher-factor tilt. That's a legit multi-stage workflow built from one-time nodes.
• Latent-space compatibility still applies - the reference should share the sampling model's VAE family.
• If the node appears to do nothing, check factor - anything near 0 is a no-op by definition.

Install

Same repo, same routine:

cd ComfyUI/custom_nodes
git clone https://github.com/RomanKuschanow/ComfyUI-Advanced-Latent-Control

Restart, or install "Advanced Latent Control" via ComfyUI Manager. No models, no pip extras.

The honest take

This is the one-time node people overlook, because a single interpolation is barely visible on its own. That's a feature: it's the safest way to add reference influence without wrecking the prompt's intent. If your windowed blends always feel too strong, this node is the fix - one tilt, then let the sampler do the rest.