Merge ratio

Model A
Model B

Formula

Block weights (MBW)

base_alpha: 0.50
Weight string

A planning helper — do the actual merge in sd-webui-supermerger / Merge Block Weighted. Paste the string into the MBW weights field; set base_alpha separately. Block counts follow the common convention (SD1.5 = 25, SDXL = 19).

A planning helper for the "ratio" and "block weights" you set when merging Stable Diffusion checkpoints. The actual merge happens in sd-webui-supermerger or the Merge Block Weighted extension — this tool just helps you decide the numbers. The first panel, Merge ratio, lets you pick a method — Weighted sum (A and B) or Add difference (A, B and C) — and drag the alpha (M) to see each model's effective contribution in the result. Weighted sum is result = A × (1−α) + B × α, so α=0.3 means 70% A and 30% B. Add difference is result = A + (B − C) × α, adding α times the B−C delta onto A (A is the base = 1, B is +α, C is −α). The formula is shown too, so you don't mix up which model dominates. The second panel, Block weights (MBW), builds the weight vector for merging the UNet block-by-block instead of with one flat ratio. Pick the architecture (SD1.5 = 25 blocks, SDXL = 19 blocks; base_alpha is separate), a shape (flat / ramp up GRAD_V / ramp down GRAD_A / smooth up / peak middle / peak ends) and the min/max weight, and it previews each block as a bar chart and generates a comma-separated weight string. Paste that string straight into supermerger's MBW weights field and set base_alpha separately. As a rule of thumb the shallow blocks (IN side) affect composition and large structure while the deep blocks (M / OUT side) affect texture and detail, which makes choosing a preset easier. Block counts follow the widely used convention, but they can differ slightly between extension versions, so confirm against what your extension expects. All math runs in your browser — the values you enter are never sent to any API or server.

How to use

  1. In Merge ratio, pick the method (Weighted sum / Add difference) and alpha to see each model's effective share.
  2. In Block weights, pick the architecture (SD1.5 / SDXL), a shape, and the min/max weights.
  3. Copy the generated comma-separated weight string and paste it into supermerger / MBW.

FAQ

Are the values I enter sent anywhere?

No. Both the ratios and the block weights are computed entirely in your browser. The values you enter are never sent to any API or server — it all stays on your device.

Can this tool actually merge models?

No. It's a planning helper for choosing ratios and weights. Do the real safetensors merge in sd-webui-supermerger, the Merge Block Weighted extension, or ComfyUI merge nodes, and feed them the numbers you get here.

What's the difference between Weighted sum and Add difference?

Weighted sum blends A and B by alpha (A×(1−α) + B×α). Add difference adds the B−C delta onto A (A + (B−C)×α). Add difference is used when you want to transfer just "what one model learned over another" into a base model.

What are block weights (MBW) good for?

They let you merge the UNet with a different ratio per block instead of one flat value. Shallow blocks influence composition and structure, deep blocks influence texture and detail — so you can, say, pull in a style while keeping the original composition.

Are the block counts (25 / 19) exact?

They follow the common convention of SD1.5 = 25 and SDXL = 19 (with base_alpha separate in both). Handling can differ slightly between extension versions, so confirm against the number of weights your target extension asks for.