Preview Bridge ⛳
Decode and preview in one node, without losing the image
- samples
- vae
- images
Every workflow ends with the same little dance: KSampler's latent goes into a VAEDecode, and the decoded image goes into a PreviewImage. Preview Bridge ⛳ collapses those two nodes into one. Feed it the latent and the VAE and it decodes and previews in a single step - you never wire up a separate decode node before looking at the result. It's a small quality-of-life win, but in a graph with five samplers it removes a lot of spaghetti.
The part that makes it worth it over just memorizing the VAEDecode → PreviewImage pattern: a plain PreviewImage is a dead end that swallows its input. This node is an output node and it returns the decoded image, so you can keep wiring the result onward - to a save node, an upscaler, a face detailer, whatever. You lose nothing by merging the pair; you just saved yourself two clicks of drag-and-drop per sampler.
How it works
The mechanism is exactly what you'd expect: it calls vae.decode() on the latent's samples, flattens 5D video-style batches down to normal 4D image batches, saves each frame to ComfyUI's temp directory as a PNG, and returns the standard UI payload so the preview renders in the node. Temp files get cleaned up by ComfyUI, so it doesn't clutter your output folder.
Inputs and outputs
samples(LATENT) - "The latent to be decoded." Comes straight off the KSampler.vae(VAE) - "The VAE model used for decoding the latent." Wire it from your checkpoint loader, or a dedicated VAE loader.- Output:
images(IMAGE) - the decoded image(s), pass-through for whatever comes next.
Install
Part of the iTools pack. ComfyUI Manager: search ComfyUI-iTools, install, restart, or clone it:
cd ComfyUI/custom_nodes
git clone https://github.com/MohammadAboulEla/ComfyUI-iTools
No requirements.txt, no models to fetch. It shows up under the iTools category. Unlike some siblings (Prompt Record is explicitly unsupported on Node 2.0 Beta), this one isn't on the pack's incompatibility list, so it should work on both node systems.
Where people get burned
The classic failure isn't the node, it's the VAE. If your preview comes out grey, washed out, or desaturated, the VAE you're feeding doesn't match the checkpoint - that's the #1 community-reported cause of flat, colorless output. Some checkpoints bake their VAE in; others expect a separate one. A mismatched VAE gives you noise or flat color, not a subtly-wrong image, so the symptom is loud. Feed it the right VAE and the preview is just a decode, nothing more. Also remember: because it writes temp PNGs, it's a preview node, not a file-keeping one - if you want the generation to stick around in your output folder, still run the images through a SaveImage at the end.
Inputs (2)
| Name | Type | Default | Description |
|---|---|---|---|
| samples | LATENT | The latent to be decoded. | |
| vae | VAE | The VAE model used for decoding the latent. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| images | IMAGE | — |