PyramidFlow Latent Preview

This is the smallest node in the pack and it's also the one that'll save you the most time. PyramidFlow Latent Preview takes raw latents and shows you roughly what they look like without running the VAE. That's the whole trick, and it's a genuinely useful one.

A VAE decode of a video latent is not cheap - PyramidFlow's VAE is a full causal video autoencoder, and decoding a 5-second clip eats real VRAM and real seconds. When you're iterating on prompts or steps, you don't want the decode in the loop. This node does the classic latent-to-RGB approximation instead: it applies a hardcoded set of 16 latent RGB factors (the same math ComfyUI uses for its fast latent previews) to turn the 16-channel latents into RGB images directly. No model runs, no tiling dance, just matrix math.

How it works

• Input: samples (LATENT) - plug in the output of PyramidFlow Sampler directly, mid-workflow or even without a decode at all.
• Outputs: images (IMAGE) and latent_rgb_factors (STRING).

The node clones the latent, applies the image and video shift/scale factors the model needs, then walks every temporal frame, does a linear transform through the RGB factors, and min-max normalizes the result. What you get out is a stack of frames that looks like a heavily-compressed, slightly-off-color draft of your video. It's a preview, not a render - the colors are shifted and detail is mushy, but composition, motion, and subject are all readable at a glance.

The latent_rgb_factors output is just the hardcoded factor matrix as a string, echoed out for the curious. You can ignore it. There's no real use for it downstream unless you want to replicate the math yourself.

Where it earns its place

• As a fast any checkpoint during a sampler run, so you can interrupt early instead of waiting for a full decode of something you already hate.
• Hanging off the sampler output in a side branch to peek at what you're about to decode, before committing the expensive VAE pass.
• Wiring images into a VHS_VideoCombine for a scratch export when you just want to check motion.

The one thing it can't do is replace PyramidFlow VAE Decode for final output - you'd never publish these colors. It's the rough cut, not the master.

Install

It ships with the pack, so nothing special: ComfyUI Manager → search "PyramidFlow Wrapper", or git clone https://github.com/kijai/ComfyUI-PyramidFlowWrapper into custom_nodes, restart. It's one of those rare nodes that needs no model download of its own - the factors are baked in. If the preview looks like noise, you've almost certainly fed it latents that don't belong to this model, since the shift/scale constants are specific to PyramidFlow's latent space. On the bright side, you can preview a latent and never once touch the VAE, which makes this a favorite for budget-GPU tinkerers squeezing every free megabyte.