Image Normalize

Image Normalize is the boring node that quietly decides whether everything after it works. The Timm Backbone Image Encode node does not touch your pixels before running the model - it expects an already-preprocessed image, and this is the node that does that preprocessing. Skip it and feed raw ComfyUI pixels straight into a backbone, and you'll get feature vectors that look like they were drawn from a random number generator. The model isn't broken; you just fed it pixels in a range it was never trained on.

The mechanism is unglamorous but worth understanding. ComfyUI hands you an IMAGE tensor in float32, with pixel values in [0, 1]. Vision backbones in timm were trained with a specific normalization: subtract a per-channel mean, divide by a per-channel std. This node does exactly that via torchvision's normalize, on each of the three channels, then hands back an IMAGE in the same [b, h, w, c] layout it received. Nothing else happens. There's no resizing here, no channel flipping - just the math.

The two inputs that matter

• mean and std - both strings, comma-separated, one value per channel. The default is 0.5, 0.5, 0.5, which maps [0,1] pixels to roughly [-1, 1]. That's a generic default, and here's the trap: it's usually wrong. What you set depends on which backbone you loaded:
  • Classic ImageNet-trained timm models (ResNets, EfficientNets, vanilla ViTs) want the ImageNet stats: mean 0.485, 0.456, 0.406 / std 0.229, 0.224, 0.225.
  • CLIP-family backbones - like the loader's default timm/vit_huge_patch14_clip_224.laion2b - want the OpenCLIP stats instead: mean 0.48145, 0.45783, 0.40821 / std 0.26863, 0.26130, 0.27578.

The model card on Hugging Face for whichever checkpoint you chose will state its normalization. Copy those numbers in.

Where people get burned

• Feeding the output into a normal img2img or VAE path. This node outputs normalized values, and the name "Image Normalize" is easy to misread as a generic preprocessor. If you plug it into anything that isn't a timm encode, the [-1,1]-ish values will look washed out or broken. This node's output belongs on the input of Timm Backbone Image Encode and pretty much nowhere else.
• Not resizing first. Normalizing doesn't fix resolution. Whatever you feed must be the backbone's training size (the 224 in ...clip_224), so put a resize node upstream before this one.
• Typo'd the mean/std strings. They're parsed by splitting on commas, so 0.485,0.456,0.406 works but a stray space-in-a-value or an extra comma will throw an error or silently produce garbage. Keep it simple: 0.485, 0.456, 0.406.

Install is the same as the rest of the pack - ComfyUI Manager search timm backbone, or clone https://github.com/p1atdev/comfyui-timm-backbone into custom_nodes and pip install -r requirements.txt (just timm). Then put it between your resize and the encode node, set the numbers your model card actually lists, and stop blaming the backbone.