Model Analyzer (Tensor Prism)

Merging blind is how you end up with a model that "feels off" and no idea why. Model Analyzer (Tensor Prism) is the pack's attempt to fix that: it takes a loaded checkpoint, pokes through its architecture and weight statistics, and hands back a plain-text report plus a JSON blob describing what's inside - block structure, layer counts, parameter magnitudes, sparsity, memory usage. And if you connect a second model to compare_with, it compares the two and suggests whether they're mergeable.

It's not a merge node; it produces no MODEL at all. It's reconnaissance, and it's the node this pack tells you to run first. The README's entire workflow philosophy is "analyze before merging," and this is the tool that makes that advice executable.

How it works

Connect a model, pick analysis_depth - Quick, Standard, or Deep. Deep takes longer and reports more; Quick is for routine checks. Then choose which sections you want via the toggles: show_architecture (block structure, layer counts, component breakdown), show_layer_stats (parameter magnitudes, sparsity, distribution metrics), show_memory_usage (memory profile by component).

The optional compare_with input is where it gets useful for merging: connect a second MODEL and the analyzer reports similarity scores and merge compatibility between the two. That's the pre-merge sanity check - merging models of the same architecture (SDXL with SDXL, same size) is the one hard compatibility rule, and this surfaces when you'd be violating it.

Two outputs, both STRING: analysis_report - the human-readable rundown with recommendations, and json_data - the same info structured for external processing. The node is flagged as an output node, so its job is to be read, not wired onward.

How to use it

• New merge project: analyze both models first, read the recommendations, then decide the merge method. The README's example workflow is exactly this: Load A → Analyze → decide method → Load B → merge.
• Two models look "same-ish" but you're not sure: use compare_with and read the compatibility verdict before spending an hour merging.
• Deep analysis when the merge matters; Quick when you're just re-checking a known model.

Installing it

Part of ComfyUI-Tensor-Prism-Node-Pack. ComfyUI Manager → search "Tensor Prism" → Install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/AstrionX/ComfyUI-Tensor-Prism-Node-Pack

No model downloads; deps are torch, numpy, psutil. Ignore the typo'd clone URL in the README.

The honest caveat

Two things to know. First, analysis runs on CPU by design (so it doesn't fight your GPU for VRAM), which means it needs system RAM - the pack's own troubleshooting notes that analysis failures are usually an out-of-memory problem, and deep analysis of a 6.5GB SDXL checkpoint on a 8GB-RAM machine can get tight. Second, this is a young first-author pack and the "smart recommendations" are heuristics, not a reliable oracle - treat its compatibility advice as a strong hint and still do the eyeball test. But as a "what's actually inside this model" tool, it's genuinely handy.