Apply Merge Recipe (Tensor Prism)

This node is the second half of a two-node idea: Analyze Model Weights computes a per-block recipe for blending two models, and Apply Merge Recipe executes it. On its own it's pointless - but as the executor of an analyzed recipe, it's the part that actually produces the merged MODEL you'd put in a KSampler.

Think of it as a recipe reader for merges. Where a hand-tuned merge is you setting one ratio and hoping the rest follows, this applies a per-block set of weights that the analysis node computed. And because strength is exposed, you can apply the same recipe at different intensities without re-running the analysis - which is the feature that makes the pair genuinely useful for iterating.

How it works

You connect model_a and model_b, plus the merge_recipe output (type MERGE_RECIPE) from an Analyze Model Weights node. These three inputs are the whole point: the recipe carries the per-block ratios, and this node carries them out.

Two controls shape the application:

• merge_method - weighted_sum (a straight weighted average per block) or add_difference (adds a scaled "B minus A" delta on top of the base). Add-difference is the more expressive of the two - it's how you graft one model's character onto another - but weighted-sum is the safer default.
• strength (default 1.0, range 0–2) - scales the whole recipe's influence. 0.5 applies the analysis at half strength; 2.0 pushes it past what the analysis suggested. This is your "the analysis is too aggressive, soften it" dial.

Optional toggles: fp16_mode (on) for memory, verbose (on) for console detail.

The single output is merged_model (MODEL), which you wire into your KSampler just like any other model.

The workflow it belongs to

[Model A] ──┐
            → [Analyze Model Weights] → [merge_recipe] → [Apply Merge Recipe] → [KSampler]
[Model B] ──┘                                              ↑
                                              [Model A]    [Model B]

The pattern to actually use: run the analysis once, then feed the recipe in at strength 0.5, 1.0 and 1.5 in three separate renders. Because analysis is the expensive step and application is cheap, you get a whole sweep of merge intensities from one analysis pass. The README also suggests saving recipes that work - they're reproducible merge plans.

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 files needed; dependencies are torch, numpy and psutil, all already in ComfyUI. The README's clone URL is typo'd - use Manager or the URL above.

The honest caveat

The whole Analyze → Apply pipeline is the kind of thing that sounds more rigorous than it currently is. It's a reasonable statistical heuristic from a first-time, openly "vibe-coded" pack - there's no large community track record behind it yet, so treat the recipes as smart starting points, not settled science. If strength 1.0 produces something off, that's more likely the analysis than your workflow; drop strength first, re-run the analysis second.