Merge Custom-Trained Models

What Is Model Merging?

Model merging is a powerful feature in Scenario that lets you seamlessly combine multiple custom-trained models into a single Multi-LoRA (also known as LoRA composition). You can merge anywhere from 2 to 5 models, allowing for flexible combinations like:

Multiple style models (e.g., “Isometric Buildings” with “3D Rendered Style”)
Style model(s) + character model(s)
Multiple character models

Why Merge Models?

Merging models gives you more precision and creative control compared to using single LoRAs. Instead of relying on trial and error with prompts, you can blend multiple influences from different models that were previously separate. With adjustable sliders, you can fine-tune each component just like tweaking ingredients in a recipe, ensuring the perfect balance for your final output.

Scenario makes it easy to test and compare different combinations step by step, so you can preview options before deciding on the best one. Here’s where model merging becomes very powerful:

Merging Different Style Models.
Combining multiple style models, such as 3D Assets and Isometric Buildings, creates a new, cohesive aesthetic. Instead of trying to force multiple styles into a single model, each style can be developed independently and then merged. Up to five style models can be blended, though 2–3 is generally recommended for optimal results. Adjusting the influence of each style ensures the right balance, providing access to a much larger range of stylistic possibilities.
Merging a Style Model with a Character Model
A style model (e.g., “chibi“) can be merged with a character model (e.g., “the sorceress“) to generate that specific character in that specific style. This method is useful for game assets, marketing creatives, and storytelling, ensuring the character remains consistent while fitting a distinct visual aesthetic
Merging Multiple Character Models
Merging character models allows for the creation of multi-character scenes without blending traits. For example, merging a “Red-Haired Superheroine” model with a “Green-Skinned Zombie” model ensures that each character maintains their defining features while interacting naturally in the same composition. This approach is ideal for game development, comics, animation, and other storytelling applications where maintaining individual character identity is crucial.

Model merging can also be applied to other options and subjects like environments, buildings, props, vehicles, and other assets, expanding creative possibilities while maintaining consistency across designs.

How to Merge Models in a Few Steps

Step 1. Navigate to the Composition Interface

Go to the “Model“ section and click "New” >> “Start Composing”. You can also directly load a model from its page by clicking on “Use this model” >> “Compose”.

Step 2. Select Your Models

Select two or more models (“single LoRA“) to merge into a new Multi-LoRA. Only models trained on the same base model (e.g., Flux-only or SDXL-only) can be merged. You can choose the base model from the panel on the left. Otherwise, once a first model is selected, only compatible models with the same base will be displayed.

Before selecting, review each model’s page, description, pinned images, and details tab to ensure it's the right choice. Use the categories (Style, Subject, etc.) to navigate models efficiently.

Step 3. Adjust Model Weights

Adjust each model’s influence using the sliders, which range from 0 to 1 in 0.05 increments. The closer a value is to 1, the stronger that model’s influence will be in the mix.

Avoid setting all weights to 1, as this can create an overly rigid or unbalanced output. A better starting point is either equal but lower weights (e.g., 0.6 or 0.7 each) or selecting one primary model with higher influence (e.g., 0.8) while keeping secondary models lower (e.g., 0.3–0.6). Experiment to find the right balance, noting that weights can always be adjusted later.

Step 4. Preview Outputs from the Merged Model (Optional)

Before finalizing and saving, preview a sample output to check the merged model’s style. Click “Test” to generate a quick test image. While not all settings from the main generation interface will be available, this allows for a fast assessment of the composition. If the result feels too blended, inconsistent, or off-style, adjust the weights and test again until it looks right.

Step 5. Save Your Multi-LoRA Model

Once you're happy with the composition, save the merged model to your workspace by clicking "Save Composition." Give it a name and add a description, especially if you're collaborating with others in the same workspace. The models included in the merge will be listed in the "Details" tab for easy reference, and you can quickly display the multi-LoRAs within the model gallery by clicking on “Composition” in the filters to the left.

The model is now ready for use, maintaining a consistent, blended aesthetic.

Step 6. Generate

You can use the new Multi-LoRA model like any other in the Image Generation screen to evaluate the model’s quality and versatility - and generate more consistent images.

However, you will notice it’s now possible to adjust model weights dynamically in the "Model Components" tab (top left): tweak slider values as needed, and save new versions instantly with the “floppy disk” icon. Or reset the initial influence with the “Reset” icon.

Final Notes

When an image is generated using a Multi-LoRA, the LoRA components and their respective weights will be displayed at the bottom-right of the image page.
Merged models also work in the Scenario Canvas, which is a great option for making scenes with multiple consistent characters for instance. For more details on multi-character workflows, check out this video
All Multi-LoRA models are compatible with Scenario Live, allowing for real-time creative exploration.
You can merge texture models just like “standard 2D image models“, expanding the range of possible texture models available.