Latent add transform (one time)

OneTimeLatentAddTransform is the one-shot version of Latent add transform: it adds a reference latent to the denoised prediction exactly once, at exactly one step, then never again. If the windowed version is "carry this reference with you through the run," this one is "slam it in once and see how the sampler absorbs it."

The inputs

• latent - the reference latent to add. Any LATENT output works; it's bicubic-upscaled to match the prediction's size if dimensions differ.
• step - the step to fire on (1-based).
• multiplier - scales the result (x + latent, then times multiplier). Near 1 to start.

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

Why one injection is interesting

Because the sampler doesn't just leave the reference sitting there - it re-noises the injected prediction and builds the next steps on top of it. One add at step 5 of 20 means the reference's structure gets folded into the trajectory, and the remaining steps refine around it. That's a softer, more organic way to get a reference's influence into a generation than the continuous version, which can feel like the reference is drowning the prompt.

It's also the cheapest way to experiment with the add-transform concept at all. A single injection is easy to reason about - you can literally see what one step of reference influence does - whereas the windowed version compounds and blurs cause and effect. Start here, then graduate to the windowed Latent add transform once you understand what you're adding and when.

Tips

• Early injection (step 2–5 of 20) = structural influence, the sampler has room to rework it.
• Late injection = a texture stamp near the finish line; fewer steps remain to soften it, so it reads more literally.
• multiplier above ~1.5 in one shot can spike the values hard - the sampler usually absorbs it, but expect a contrast jump for a step or two.
• Keep the reference in a compatible latent space (same VAE family as the model you're sampling with), or the injection is just noise.

Install

Standard for the pack:

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 extra dependencies.

The bottom line

One-time add is the gentlest intro to mixing reference latents into sampling, and a great diagnostic: it shows you exactly what a single injection does before you commit to a windowed approach. Keep the multiplier modest, keep the reference in the right latent space, and you'll learn more from one run with this node than from ten runs of the windowed version.