BlehLatentAsImage

The mirror image of BlehImageAsLatent: this node rearranges an actual diffusion LATENT into something typed as an IMAGE, mainly so you can point image-space tooling at it, or just look at the raw numbers packed inside a latent tensor. Like its sibling, there's no VAE decode happening here - this is tensor reshaping, not a way to see your actual generated picture. If you want that, use VAEDecode.

Inputs that matter: latent (required LATENT). values_mode (default rescale) controls how the latent's numeric range gets mapped into the 0-1 range IMAGE tensors expect - rescale does this globally across the whole tensor, rescale_perchannel does it independently per channel instead, which can look meaningfully different since latent channels don't share the same value distribution, and clamp just clips values into range rather than remapping them, which is a different (and sometimes more honest) way to visualize what's actually there. channels_into_batch (default false) is the setting worth knowing about if you're doing this for genuine inspection rather than as a pipeline step: enabled, it produces a separate greyscale image per latent channel instead of one combined, fairly arbitrary-looking RGB-ish composite - much more useful if you actually want to see what each channel is carrying.

Output is a single IMAGE (or a batch of them, per-channel, if channels_into_batch is on).

It pairs with BlehImageAsLatent for a full round trip - reshape into LATENT, run whatever latent-space operation you needed, reshape back with this node to get a usable image out the other end. That round trip is really the main reason to use either node on its own: on its own, BlehLatentAsImage is mostly a debugging tool, since a real generated image still needs an actual VAEDecode to look right. Where it earns its keep is letting you borrow bleh's wider set of image-space filters and blend nodes for something that started life as a latent, or the reverse - apply one of bleh's latent-space blend modes to something that started as a plain image via BlehImageAsLatent, then bring it back with this node.

One real limitation worth knowing before you rely on this for anything precise: the node's own description states plainly that it does not respect masks or the kind of latent-selection metadata that nodes like LatentFromBatch attach to a latent. If your latent is carrying that metadata, this node ignores it entirely and just processes the raw tensor.

Install is the standard route for the pack: ComfyUI Manager, or git clone https://github.com/blepping/ComfyUI-bleh into custom_nodes followed by a restart. No extra dependencies.

Worth a practical note on when to actually reach for this versus something else in the pack: if what you want is to preview what a latent looks like mid-generation, bleh's own previewer - the feature most people install this pack for in the first place - is a much better tool for that job, with proper TAESD decoding rather than a raw channel dump. Use this node specifically when you need genuinely IMAGE-typed output to feed into something downstream that only accepts that type, not as your go-to way to sanity-check a latent while you work.