Weighted Sum

Weighted sum is the oldest, most boring, most used merge in the hobby: result = (1 - alpha) * model_a + alpha * model_b. Two checkpoints, one slider, and you get a blend - 50/50 of a realism base and an anime base, a touch of style model A over backbone B, whatever. In the A1111 world you'd reach for SuperMerger for this; in ComfyUI this node from the comfy-mecha pack is the modern equivalent, and it has one big advantage: it doesn't need to load both full models into VRAM at once.

How it works

Like every node in this pack, Weighted Sum Mecha Recipe doesn't merge anything itself. It builds a MECHA_RECIPE - a graph of instructions describing the merge - and the graph only gets executed when you feed it to a Mecha Merger. sd-mecha (the underlying library, also by ljleb) processes the merge one tensor key at a time, reading from the safetensors header and streaming results to disk. The author's own pitch is that you can run a weighted sum on a phone. The practical version of that: an SDXL-weighted-sum merge that used to want most of your system RAM now runs comfortably alongside a browser.

Inputs:

• a and b - two MECHA_RECIPE inputs (wire in Model Mecha Recipe nodes for your checkpoints).
• alpha - the blend. Default 0.5, so a straight 50/50. 0 gives you pure A, 1 gives you pure B. The interesting part: alpha is itself a MECHA_RECIPE, so you can feed it per-block weights (see Blocks Mecha Hyper) for merge-block weighting, or a Float recipe for a uniform value. That's how you get the "MBW" merges people obsess over.

Output is one recipe (MECHA_RECIPE), which you chain into Mecha Merger to get a MODEL, CLIP, and VAE you can save.

The merge_checkpointing toggle on the node caches the branch on CPU in fp16 once computed - handy when you're running the same expensive merge repeatedly while fiddling with other inputs. The full default-merge settings (device, dtype, buffer size, threads) live on Mecha Merger, not here.

A note on merges in general

Merging feels like a cheat code until you hit the reality that stacked weighted sums drift toward the mush between two models. Weighted sum is a great first tool, but for "add one thing to another" you usually want Add Difference (a + alpha * delta), and for combining many models the TIES-style merges in this pack exist for a reason. If your goal is just "make checkpoint C look more like B," start here; if you're chasing a specific character or style, a LoRA merge is usually cleaner.

Install

ComfyUI Manager → search mecha → install "Mecha Merge Node Pack". Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/ljleb/comfy-mecha.git
pip install -r comfy-mecha/requirements.txt

Restart after. Only dependency is sd-mecha==1.1.7; no models to download.

Gotchas

• Both inputs need to be in the same merge space (weight) for a plain blend - the node labels tell you which spaces are accepted. If you feed a delta where a weight is expected, the merge may silently produce garbage or refuse with a validation error, which is the better outcome.
• Don't forget Mecha Merger's output_dtype - default fp16 is fine for most, but if you're merging at high precision for later distillation, flip it to fp32 and set default_merge_dtype accordingly on the merger.
• The pack's whole trick is streaming; if you see the RAM spike anyway, check that total_buffer_size on the merger isn't set absurdly high.