Timm Embeds Print

Timm Embeds Print is a debugging node, and it is the most underrated thing in this pack. The whole point of the pack is that Timm Backbone Image Encode spits out a raw TENSOR that no standard ComfyUI node understands - so the first question everyone hits is "did it actually work, and what did I get?" This node answers both by printing the tensor's shape and contents to your ComfyUI console.

It's an output node, which means it's a dead end in the graph: plug your tensor in, run the workflow, and watch the terminal. You get one input, embeds, and no outputs - nothing to wire onward. (Don't worry about what it prints vs. the README calling the input "tensor"; the actual node input is embeds, and that's what ComfyUI will show you.) When the queue finishes, the console shows something like:

torch.Size([1, 768]): tensor([[ 0.0123, -0.0456, ...]], device='cuda:0', dtype=torch.float16)

That single line tells you most of what you need to know.

What to look at in the output

• The shape. This is the real payoff. [1, 768] means you got the pooler output - one global vector per image. [1, 256, 768] means hidden-state mode, 256 patch tokens for a 224px patch-14 ViT. If you expected a vector and see a 3D tensor, you're in the wrong feature_type; if you expected tokens and see 2D, same fix.
• The dtype. torch.float16 is correct and expected - the pack casts to fp16 internally. If you ever see NaN values in the printout, that's your model producing bad math (a known trap on older GPUs with weak fp16 support), not a node bug.
• The device. cuda:0 means it landed on your GPU as intended. If it says cpu, either you're running a CPU-only ComfyUI or something upstream forced a move - fine for small models, slow for the default ViT-Huge.

Why you'll actually want it

Because the pack's output type doesn't connect to anything except your own custom nodes, "does my tensor look right" is not a question you can answer visually. This node is the two-second sanity check between your encode and whatever custom node you're building. Get in the habit: normalize → encode → print, confirm the shape and no NaNs, then swap the print node for your actual consumer. It costs one run and saves an hour of "why is my custom node receiving garbage."

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. Then remember where the printout lands - it's in the terminal that launched ComfyUI, not in the browser UI, which trips up more than a few people on their first run.