Latent interpolate transform

Latent interpolate transform is the softer sibling of the pack's Latent add transform. Instead of adding a reference latent to the denoised prediction, it blends toward it - a weighted average between the current prediction and your reference latent, controlled by a factor. Think of it as a smooth steering wheel for the sampling trajectory rather than a kick.

The inputs

• latent - the reference latent to blend toward. Like its add sibling, it's bicubic-upscaled to match the prediction's size if needed.
• factor - the blend weight, 0 to 1. At 0 the transform leaves the prediction untouched; at 1 it replaces the prediction with the reference; 0.5 is a true halfway blend. The default is 0.5.
• multiplier - scales the blended result afterwards.
• start_at / stop_at - step window as fractions.
• offset_optional - gate to every Nth step.

It emits a TRANSFORM for a TSampler, TSampler Advanced, or Transform Hijack-patched sampler.

Why you'd use it

Because "blend toward a reference" is gentler than "add a reference," and gentleness is usually what you want mid-sampling. Add a strong latent and you've injected its energy wholesale; interpolate toward it at factor 0.3 and you've tilted the trajectory toward its character without letting it take over. That's the node for guiding a generation toward a style, mood, or composition that the prompt alone isn't hitting.

The factor ramp is the fun part. Because the factor is fixed per transform but the transform's window can be anything, you can do coarse stage-based control: low factor early to plant the reference's structure, none late so the model finishes on its own. If you want a ramp (progressively more reference over the run), you'd chain a couple of interpolate transforms with different factors and staggered windows via Combine transforms - a two-step version of a continuous blend.

Watch out for

• Latent-space mismatch again: the reference should come from a compatible VAE/latent space, or the blend is garbage-in-garbage-out.
• factor: 1 in a replace-style full window is just "keep pasting the reference in" - that's not a blend, that's a hostage situation. Keep the factor modest.
• The standard pack trap applies: start_at: 0, stop_at: 0 never fires.

Install

One repo, same as always:

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

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

Where it fits

If you've ever fought img2img denoise strength trying to "keep the mood but change the subject," this is the transform that does that mid-flight: encode your mood reference, blend toward it at factor 0.2–0.4 during the middle steps, and let the tail steps diverge. It's one of the more expressive nodes in the pack precisely because it's mild - and mild is a feature, not a bug.