Latent Blend

LatentBlend is the most honest node in the experimental category: it takes two latent tensors, mixes them with one slider, and hands you the result. That's the whole job. It's been in core since the earliest days - it predates the modern guidance hacks by years - and its longevity is the point. Sometimes you need a boring interpolation tool, and this is the one ComfyUI ships.

The math. Result = samples1 × blend_factor + samples2 × (1 − blend_factor). At blend_factor of 1.0 you get exactly samples1; at 0.0 you get samples2; at 0.5 it's a perfect even mix. It operates on the latents, not pixels, which matters: you're blending in the compressed latent space the model actually works in, before the VAE ever decodes anything. That's where latent interpolation lives.

Inputs and output.

• samples1 / samples2 - two LATENT tensors.
• blend_factor (default 0.5, range 0–1) - the only knob you'll actually touch.
• Output is a LATENT, so it plugs anywhere a latent goes - into a KSampler, into a LatentUpscale, whatever your graph needs.

What people actually use it for. The classic case is latent interpolation: generate two latents from two prompts (or two points in an img2img sequence) and step blend_factor from 0 to 1 to morph between them. It also shows up in img2img workflows where you want to blend a partially-denoised latent back toward the source, or in experiments where you mix the same seed's latents at different steps. There's a hidden detail in the code worth knowing: if the two latent batches differ in shape, it silently upscales samples2 to samples1's resolution (bicubic, center-cropped) before blending. That's usually not what you want - it's a convenience, not a feature - so keep both latents at the same resolution unless you deliberately want the implicit resize.

Gotchas. Because it works on raw latents, the tensors you mix need to be "compatible" in a looser sense than pixel blending: mixing latents from different models or wildly different denoise states can produce muddy results, because latent spaces aren't linear between checkpoints. And there's no mask or region support - it's a global blend across the whole tensor. If you want to blend only part of an image, that's what latent masking or ImageCompositeMasked is for; LatentBlend is the whole-canvas tool.

The experimental tag here undersells it. Unlike the guidance patchers in this category, LatentBlend is a tiny, stable utility with no model to patch and no CFG to fight - the risk of it breaking your workflow is close to zero. When you're hunting for a "mix latents" or "interpolate latents" node in the search box, this is the one ComfyUI itself offers.

Ships with ComfyUI core. No install, no models, no dependencies - double-click, search "LatentBlend," wire it in, done.