Model Comparator (Tensor Prism)

Before you spend an evening merging two checkpoints, it's worth asking one dumb question: are these two even mergeable? Model Comparator (Tensor Prism) is that question, automated. You wire in two models, and it returns a similarity score, a boolean compatible verdict, and a short text report - a compatibility screening before you invest in the actual merge.

The core rule of checkpoint merging is brutally simple: the two models must share an architecture. Merge an SDXL with an SDXL fine-tune and you're cooking; try to average an SDXL with something else and you get garbage or a crash. The KB's checkpoint panel is explicit that merging incompatible targets dilutes both parents - and the Comparator is a fast, cheap way to check you're not about to do that, plus a rough sense of how similar two compatible models are.

How it works

model_A and model_B, nothing else required - there are no optional inputs. The node compares the two models' weights and produces three outputs:

• comparison_report (STRING) - a human-readable rundown of the comparison, with merge recommendations.
• similarity_score (FLOAT) - a quantitative similarity measure, which the README frames as the basis for your merge planning. Very high similarity means the merge will be boring (and mostly redundant); low similarity means interesting but risky.
• compatible (BOOLEAN) - the yes/no verdict. This is the one that matters on day one.

The node is flagged as an output node, so its outputs are meant to be read (in a text node) or routed into workflow logic - the README explicitly suggests using the boolean for workflow decisions, e.g. skipping a merge branch when models are incompatible.

How to use it

• First thing on any merge project: run both models through this before anything else. Incompatible → stop. Compatible but nearly identical → the merge is probably a waste of time. Compatible and different → proceed.
• Pair it with Model Analyzer if you want the deeper per-block breakdown; the Comparator is the 30-second gate, the Analyzer is the full investigation.

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. The README's clone URL is typo'd - use Manager or the URL above.

The honest caveat

The similarity score is a heuristic from a young, first-time author's "vibe-coded" pack, so don't treat the exact number as a calibrated instrument - treat compatible as a real guardrail and similarity_score as a rough "how close are these" gauge. And remember the subtle version of the rule the score can't capture: even two compatible models can be philosophically opposite (a realism fine-tune vs. an anime fine-tune are both SDXL and will happily "merge" into something worse at both). The Comparator tells you if the merge is possible; it's still on you to decide whether it's wise.