Error messages

Given groups=1, weight of size

Given groups=1, weight of size

The workflow runs until a sampler or a VAE decode node, and the error report says:

## Error Details
- **Node ID:** 10
- **Node Type:** VAEDecodeTiled
- **Exception Type:** RuntimeError
- **Exception Message:** Given groups=1, weight of size [512, 16, 3, 3], expected input[1, 4, 128, 32] to have 16 channels, but got 4 channels instead

That one's from issue #6769. The numbers differ from report to report, but 50 issues in the ComfyUI repo have the same opening words as of September 26, 2026. The text comes from PyTorch, which does the actual maths, and ComfyUI passes it through unchanged.

Reading the numbers

Image models work on arrays with several channels, the way a photo has red, green and blue. The "weight of size" list belongs to the layer that failed, and its second number is how many channels that layer takes in: 16 in the example. The "expected input" list is what actually arrived, and its second number is the channel count it carried: 4.

So the message reads as "this part was built for 16-channel data and got 4-channel data." "Groups=1" is a PyTorch setting on the layer and you can ignore it.

Why it happens

Different model families use differently shaped latents, the compressed images a diffusion model works on. Mix pieces from two families and one of them gets the wrong shape. A ComfyUI team member's reply in #6769 covers the most common case: "You cannot mix SD diffusion model and FLUX VAE model. Or vice versa. You must always ensure that the VAE model you apply matches the diffusion model."

The same goes for latents passed between samplers. In #2838, a Stable Cascade workflow had the wrong latent wired into a sampler, and the team member spelled out the two shapes: "stage_c model requires 16 channel latent. stage_b model requires 4 channel latent."

Video models hit it too. A Wan 2.2 report (#9596) had a layer wanting 36 channels and getting 32. No team member answered it, but one user fixed it by switching to the Wan 2.1 VAE file, and another said that swap didn't help, so not every case has a clean explanation on record.

How to fix it

Start at the node the report names and walk backwards along its inputs. Check each loader: the checkpoint or diffusion model, the VAE, the CLIP or text encoder, any ControlNet, and anything producing a latent (an Empty Latent node, an image encode, a previous sampler). They should all be for the same family. Take extra care with a workflow someone else made, since its loaders keep whatever file was last picked in them and file names don't always say which family a file belongs to.

If every part really does match (one person in #4988 said both their checkpoint and ControlNet were SD 1.5 and still got it), disable custom nodes and run it again, which is the first step ComfyUI's own troubleshooting guide asks for.