VAE Decode All Steps

This node is the finishing move for the pack's main trick. Its sibling, Sampler (Custom Advanced All Steps), hands you a batch of latents - one per denoising step. This node, VAE Decode All Steps, takes that batch and turns it into one image per step, so you can finally look at what the model was thinking the whole way down.

On its own it's honestly not doing anything the core VAE Decode can't do - batches of latents decode fine through the built-in node. The reason this one exists is pairing and predictability: it's the sibling explicitly built for the all_steps_latents output, and it loops over the batch one latent at a time instead of blasting the whole thing through at once, which keeps memory behavior predictable on long runs.

How it works

Two inputs, one output - this is about as simple as a node gets:

• samples (LATENT) - the all_steps_latents output from SamplerCustomAdvancedAllSteps
• vae (VAE) - your VAELoader

It slices the incoming batch into individual latents, decodes each one through the VAE on its own, then concatenates the decoded images back into a single images (IMAGE) output. One image per step, in order, all in one batch. From there it's the usual: PreviewImage gives you a grid of the whole denoising trajectory, SaveImage writes them all out.

Why you'd reach for it

Say you're tuning a distilled turbo model and you keep wondering "is it actually done at step 4, or does step 6 still change things?" Feed all_steps_latents in here and you get side-by-side proof instead of vibes. Or you're debugging a burnt, over-cooked render - you can spot the exact step where the image went wrong. Or you just want a clean denoising-progress animation. Same underlying idea as the sampler node: stop trusting that one final frame and actually watch the trajectory.

Installing it

Same pack, same trivial install. Via ComfyUI Manager, search ComfyUi-DisplayAllSteps, or:

cd ComfyUI/custom_nodes
git clone https://github.com/BigStationW/ComfyUi-DisplayAllSteps

Restart ComfyUI. No requirements.txt, no model files, no extra dependencies - everything it needs ships with ComfyUI itself.

The one real gotcha

Decode time scales with step count. This node is N VAE decodes, one per frame. A 4–8 step turbo grid is basically instant, but a 30-step SDXL run means 30 full-res decodes plus a grid that's 30 images wide - you'll wait, and your RAM will notice. If all you want is a few checkpoints along the way, this isn't the tool - but if you want every step, this is exactly the tool. Keep step counts sane and it's a genuinely great little debugging utility.