CogVideo LatentPreview

This is a cheap way to look at a latent without paying for a full VAE decode. A proper decode (via CogVideo Decode) is accurate but slow and VRAM-hungry; this node instead does a quick linear approximation - projecting the latent straight to RGB using a set of tunable factors - to give you a rough, fast preview of what's in there. It's a diagnostic tool, not a final-output tool. Use it to sanity-check that a generation is going somewhere reasonable before you commit to the real decode.

A small oddity worth mentioning: this node's category is listed as PyramidFlowWrapper rather than CogVideoWrapper. That's a leftover from shared code between kijai's wrappers - it still lives in and works with the CogVideoX pack, it just carries a category label from its origins. Don't read anything into it.

The inputs that matter

The two you care about are samples (the LATENT to preview) and, honestly, not much else - the rest is color-approximation tuning:

• min_val (-0.15) / max_val (0.15) - the value range the approximation maps to color. These clamp the projection; the defaults are tuned for CogVideoX latents and you rarely need to touch them.
• r_bias, g_bias, b_bias (all 0) - per-channel color nudges. If your quick preview has an obvious color cast, these let you correct it, but since it's only an approximation, it's usually not worth fussing over.
• seed - present for consistency with the sampler's seed handling.

Two outputs: images (an IMAGE batch - the preview frames) and latent_rgb_factors (a STRING holding the projection coefficients, mostly useful if you're deriving or reusing a set of factors). The images output goes to a preview or save node just like any image.

How to install it

Part of the CogVideoX wrapper:

• ComfyUI Manager - search ComfyUI-CogVideoXWrapper, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-CogVideoXWrapper, then pip install -r ComfyUI-CogVideoXWrapper/requirements.txt, restart.

Common issues & troubleshooting

The preview looks worse than my final output. That's expected and fine. This is a linear approximation of the latent, not a real decode - colors will be off, detail will be crude, and it's meant to be. If you want an accurate picture, use CogVideo Decode. Judge the composition and motion, not the fidelity.

The colors are way off. Adjust min_val/max_val to reframe the value range, or the r/g/b_bias to correct a cast. But temper expectations - no amount of tuning makes an approximation match a true VAE decode, so don't chase perfection here.

Do I need this at all? For a finished workflow, no. It's a convenience for iterating quickly or eyeballing intermediate latents. Plenty of CogVideoX graphs never use it and just decode at the end.