What does the alpha coefficient control?

Alpha is the blend ratio. In a weighted-sum merge, alpha is the fraction of model B mixed in — alpha 0 is pure model A and alpha 1 is pure model B. Sigmoid methods reshape that curve to favour one model more strongly in the middle range.

When should I use add-difference?

Add-difference transfers what makes model B special relative to a base model C. The formula is A + (B - C) * alpha. Use it to graft a fine-tune's style onto another checkpoint while keeping that checkpoint's structure.

Do merged models need re-training?

No. Merging is a weighted average of weights and happens in seconds with no GPU training. That is why it is so popular for quickly combining the strengths of two community checkpoints.

Why is my merge producing washed-out or broken images?

Extreme alphas or merging incompatible architectures (for example an SD 1.5 model with an SDXL model) breaks the weights. Keep alpha in a sensible range, only merge same-architecture checkpoints and test at a few alpha values.

What is the SD Model Merge Recipe Builder?

Build Stable Diffusion model merge recipes. Pick two or three checkpoints, choose a merge method (weighted sum, sigmoid, inverse-sigmoid, add-difference) and set coefficients to produce a ready-to-run Automatic1111 merge command. It runs free in your browser on Gera Tools, with nothing uploaded.

SD Model Merge Recipe Builder

Name: SD Model Merge Recipe Builder
Creator: Gera Tools
License: https://creativecommons.org/licenses/by/4.0/

Model merge recipe builder

Merging Stable Diffusion checkpoints is a weighted average of their weights — no training required. The result inherits traits from each parent in proportion to the alpha coefficient and the merge method you choose. This builder assembles a clean, reproducible recipe and the exact Automatic1111 Checkpoint Merger settings so you can recreate or share the blend precisely.

How it works

Three methods take two models: weighted sum (A*(1-a) + B*a), sigmoid and inverse-sigmoid, which apply a non-linear blend curve so one model dominates more in the middle of the range. The add-difference method takes three models and computes A + (B - C) * a, transferring the difference between B and a base model C onto A. The tool validates that you supplied model C when add-difference is selected and clamps coefficients to the valid 0-1 range.

What each merge method does

Weighted sum is the most common starting point. At alpha 0.5 you get equal contributions from both models. Alpha 0.3 gives 70% of A plus 30% of B, leaning toward A’s character. Alpha 0.7 tilts it toward B. This is linear — the blend curve is a straight line between the two models.

Sigmoid and inverse-sigmoid reshape that blend curve. Sigmoid weights the midpoint toward model A, so you need a higher alpha to feel model B’s influence. Inverse-sigmoid does the opposite, giving model B more influence earlier. These matter when a straight weighted sum gives an unsatisfying middle state rather than cleanly taking what you want from each model.

Add-difference (A + (B - C) × alpha) is for surgical style transfers. Model C is the base that both A and B were trained on or fine-tuned from. The “difference” B - C represents everything that makes B distinctive relative to that base. Adding it to A transplants B’s aesthetic character into A’s structure. For example, if A is a high-quality photorealistic base, B is an anime fine-tune, and C is the same base checkpoint both were derived from, add-difference at alpha 0.5 adds half of the anime characteristics to A.

Example recipe

For example, merging a landscape-focused checkpoint with a portrait-focused one using weighted sum at alpha 0.45 would give a blend that leans slightly toward the landscape model while retaining meaningful portrait quality. The builder records this as:

Method: Weighted Sum
Model A: landscape-v2.safetensors
Model B: portrait-realistic.safetensors
Alpha: 0.45
Output: landscape-portrait-blend-045.safetensors

Run the same prompt and seed at alpha 0.30, 0.45, and 0.60 to compare before committing to a final value — the optimal blend point is rarely obvious without testing.

Tips for clean merges

Only merge same-architecture models. SD 1.5 with SD 1.5, SDXL with SDXL — never mix the two; mismatched tensor shapes produce garbage output or an error.
Sweep alpha before finalizing. Test at 0.3, 0.5 and 0.7 with identical seeds to find the sweet spot visually, rather than committing to a single value blind.
Use add-difference for style transfer. It is the cleanest way to graft a fine-tune’s aesthetic onto a stronger base without pulling in its training artifacts.
Bake the VAE separately. Merging does not change the VAE; attach a known-good VAE at inference time for consistent colour and sharpness.
Save the recipe. The builder outputs the settings you can paste into Automatic1111’s Checkpoint Merger tab so you or a collaborator can reproduce the exact blend later.