BlehModelProcessLatentOut

Same story as its sibling BlehModelProcessLatentIn, mirrored: every model architecture has its own internal scaling convention, and just as it scales a latent up to its expected range before denoising, it also has to scale the result back down to a "normal" range once denoising is done - that's what feeds cleanly into a VAE decode or gets saved out. Normally you never see this happen; the standard sampler and decode nodes call it for you as part of the normal pipeline. This node makes that hidden process_latent_out call something you can invoke directly.

The pack's own framing is blunt about who needs this: "generally only needed if you're doing something that bypasses the normal latent output mechanisms." If your workflow ends at a regular VAEDecode fed by a regular sampler, this scaling already happened for you and you have no reason to touch this node.

When it's actually useful

This is the counterpart of BlehModelProcessLatentIn for custom sampling setups: if you've built your own denoising loop, or you're extracting a latent partway through a custom pipeline and need it in the model's "final" range before doing something with it - decoding it manually, exporting it, feeding it into another tool - that scaling doesn't happen automatically outside the normal rails. This node lets you apply it explicitly, on demand, at exactly the point in your custom graph where you need it.

If you paired BlehModelProcessLatentIn to get a raw latent into a model's expected range, this is the node that undoes that on the way back out.

Inputs and outputs

• model - the MODEL whose output-latent convention you want applied. Needs to match the architecture the latent actually came from - this scaling is architecture-specific, not generic.
• latent - the LATENT to process.
• Output: a single LATENT, scaled into that model's expected output range.

Installing it

• ComfyUI Manager - search "ComfyUI-bleh", install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/blepping/ComfyUI-bleh, restart ComfyUI.

No extra dependencies or model downloads for this node.

Common issues

Colors or brightness look off after decoding a latent you processed with this node. If your latent had already gone through the model's normal output path (say, it came straight out of a regular KSampler), running it through this node again double-applies the scaling. Only reach for this on latents that genuinely bypassed the normal output handling.

Not sure if you need this at all. If your pipeline is a standard sampler feeding a standard VAEDecode, you don't - that path already does this for you. This node exists for custom sampling graphs that need the same math available as an explicit, callable step.

Paired it with BlehModelProcessLatentIn and something's still wrong. Check that both nodes are pointed at the same model - mixing latent-in scaling from one architecture with latent-out scaling from another will produce a latent that isn't correctly calibrated for either.