🖼️ TA Latent Preview

Latents are invisible, which makes debugging annoying. TA Latent Preview is a tiny node that fixes exactly that: feed it a latent, and it projects the tensor into a rough RGB image you can actually look at - no VAE decoder, no TAESD model, nothing extra to load.

It's one of those nodes that sounds trivial until you need it. After a KSampler, the standard way to see your work is VAE Decode → Preview Image, but that costs a full VAE pass and pins more VRAM. This node instead applies the same latent2rgb projection matrix ComfyUI's own progress previews use - a fixed linear map from latent channels into RGB - and renders that. It's approximate, obviously: it looks like a washed-out, slightly psychedelic version of your image, not the final decode. But it's enough to tell "composition is completely wrong" from "working as intended", and it runs anywhere, because it's just matrix math on the CPU if need be.

The single required input is latent, straight from a sampler. The output is an IMAGE tensor, so it isn't just for eyeballing - you can pass the projection downstream to any image node if you want to, say, check it, resize it, or feed it into a comparison node. It's marked as an output node, so the preview also renders inside the UI like ComfyUI's PreviewImage, writing temp PNGs to the standard place the frontend expects. Both the 4-channel latents from SD/SDXL and the 16-channel ones from SD3/Flux-family models are handled by slicing the factor matrix to match the channel count, so it's safe to drop into most samplers.

Where it earns its keep: checking a latent between sampling stages in a multi-pass workflow, sanity-checking an img2img input before you commit to a run, or previewing when you're out of VRAM and can't afford a real decode. Think of it as the debug view, not the final image.

Install. Part of the TA pack, which installs with zero Python dependencies:

cd ComfyUI/custom_nodes
git clone https://github.com/tmode-1960/TA-ComfyUI-Nodes-Pack

or via ComfyUI Manager (search "TA ComfyUI Nodes Pack"), then restart. You'll find it under TA Nodes/Sampling.

Gotchas. The big one is expectations: this is a latent projection, not a decode. If you show the output to someone and call it a final image, you'll have some explaining to do. It's also per-frame, so a big batch means a pile of small preview PNGs in ComfyUI's temp directory - harmless, but it adds up on long runs. And one pack-level note: if you're loading an older TA v1 workflow, the nodes were renamed and rewired for v2.x, so rebuild rather than fight it.