Model Info (Oli)

Model Info (Oli) is a debugging node for model pipelines. Connect anything - a MODEL, a CLIP, a VAE, whatever - and it tells you what it actually is: the Python class name, the hidden dimension, the parameter count, whether it's a standard or GGUF-quantized load, and the title of the node that produced it. It's the "wait, what did I actually connect here" node, and it earns a slot in any workflow you're trying to untangle.

How it works

It's introspective rather than magical. For a MODEL input it unwraps the ComfyUI ModelPatcher to reach the underlying diffusion model, reads its class name and model_type, then hunts for a hidden dimension across the usual attribute names (hidden_size, dim, embed_dim, hidden_dim, d_model, width, and friends). It also counts parameters - that's where a "1.3B vs 14B" answer comes from - and marks the format as gguf or standard, so a mislabeled quantized load shows up immediately. CLIP and VAE inputs get the same class/params treatment, minus the dimension.

The label output walks the workflow graph to find the title of the node that produced your model. Name your loaders sensibly and label becomes a clean routing signal.

Inputs and outputs

Just one input: model (any type, optional). Outputs:

• class_name - the Python class of the underlying model (WAN21, Flux, etc.)
• dim - hidden dimension (0 for CLIP/VAE, which don't expose one the same way)
• label - the upstream node's title

It's marked as an output node, so after each run it also prints its findings on the canvas - you get the readout without adding a text-display node to the graph.

Install & gotchas

cd ComfyUI/custom_nodes
git clone https://github.com/magicoli/comfyui-oli-prompt-tools

…or ComfyUI Manager → Oli Prompt Tools, then restart. No dependencies.

Caveats: dim is best-effort - it reads whatever attribute names the model happens to expose, and if none fit the pattern it returns 0. Don't read a missing dim as a broken install. And label needs the workflow graph info, which the UI always has; in pure API execution it can come back empty. It's a diagnostic, so expect it to feel a little plain - that's the job.