Icy LatentBlend

IcyLatentBlend is the ComfyUI-IcyHider wrapper of core LatentBlend - same node, renamed class so the privacy extension can hide it. Install the pack or use the core node; there's no functional difference.

The node blends two latents - the model's compressed image representation - by linearly interpolating between them. samples1 and samples2 (both LATENT) are mixed according to blend_factor (0–1, default 0.5): 0 gives you all of samples1, 1 gives you all of samples2, 0.5 is a 50/50 mix. Output is a single LATENT.

Why blend in latent space at all

Because a latent blend is not a pixel crossfade. When you interpolate in latent space, you're mixing the semantic features the model has already organized - structure, composition, style - rather than just averaging colors. This is the same family of trick that makes "morph" workflows work: two images sampled to latents, blended at 0.5, and you get a genuinely hybrid image that shares the DNA of both, with VAE decoding after. The KB's concepts material frames latents as the model's working space, and that's exactly why the blend behaves so much more usefully than a pixel dissolve.

Where people actually use it:

• Concept mixing. Sample two prompts, blend their latents, decode - a cheap way to explore the space between two ideas without retraining anything.
• Style transfer-lite. Blend a generated image's latent with a reference's.
• Smoothing a transition in multi-image or animation-ish workflows.

The gotchas

• Latents must be compatible. Both inputs need matching dimensions and the same VAE origin. Blending a 512×512 latent with a 768×768 one (or two latents from different VAEs) errors out or produces nonsense. Normalize their sizes before blending.
• It's not a picker. blend_factor does a straight linear mix, so at 0.5 you get a blend, not a choice between the two. If you wanted to place one inside the other, that's LatentComposite, not this.
• The wrapper is cosmetic. As with every node in this pack, "Icy" is just the hiding extension's naming - the interpolation math is core ComfyUI's.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/icekiub-ai/ComfyUI-IcyHider

Restart, or use ComfyUI Manager (search "IcyHider"). No pip dependencies, no models.

Common issues

• Missing node in shared workflows - install the pack or swap in core LatentBlend.
• Shape mismatch errors. The two latents don't line up. Resize them to match first (via their image path or a latent-resize node) and try again.
• "The result looks mushy." A 50/50 latent mix often does at first - that's a real property of the technique, not a bug. Tilt blend_factor toward one side (0.7/0.3) or blend earlier in the schedule (before full sampling) for cleaner results.