Implementation:AUTOMATIC1111 Stable diffusion webui Merge output configuration

Knowledge Sources	stable-diffusion-webui
Domains	Model Precision, Checkpoint Merging, Model Optimization
Last Updated	2026-02-08 00:00 GMT

Overview

Concrete tool for configuring precision, VAE baking, metadata, and weight filtering of merged model output provided by stable-diffusion-webui.

Description

The merge output configuration consists of several utility functions and code paths within modules/extras.py that control how merged model weights are post-processed before saving:

• to_half(tensor, enable): Conditionally converts a float32 tensor to float16. Returns the tensor unchanged if enable is False or if the tensor is not float32 (e.g., integer tensors are never converted).
• read_metadata(primary, secondary, tertiary): Collects metadata dictionaries from up to three source checkpoint models and merges them into a single JSON string. This is used to populate the metadata preview in the UI.
• checkpoint_dict_skip_on_merge: A list of state dict keys that are always skipped during merging (e.g., cond_stage_model.transformer.text_model.embeddings.position_ids) because they are not weight tensors and should not be interpolated.
• discard_weights: A regex pattern applied post-merge to remove unwanted keys from the output state dictionary.
• bake_in_vae: Replaces VAE weights in the merged model with those from a specified external VAE file.

Usage

These configuration options are passed as parameters to run_modelmerger and applied during the merge pipeline. The to_half function is applied per-tensor during the merge loop, while VAE baking and weight discarding happen after all interpolation is complete.

Code Reference

Source Location

• Repository: stable-diffusion-webui
• File: modules/extras.py
• Lines: L65 (checkpoint_dict_skip_on_merge), L68-72 (to_half), L75-85 (read_metadata), L221-242 (VAE baking and weight discarding within run_modelmerger)

Signature

checkpoint_dict_skip_on_merge = ["cond_stage_model.transformer.text_model.embeddings.position_ids"]

def to_half(tensor, enable):
    if enable and tensor.dtype == torch.float:
        return tensor.half()
    return tensor

def read_metadata(primary_model_name, secondary_model_name, tertiary_model_name):
    metadata = {}
    for checkpoint_name in [primary_model_name, secondary_model_name, tertiary_model_name]:
        checkpoint_info = sd_models.checkpoints_list.get(checkpoint_name, None)
        if checkpoint_info is None:
            continue
        metadata.update(checkpoint_info.metadata)
    return json.dumps(metadata, indent=4, ensure_ascii=False)

Import

from modules.extras import to_half, read_metadata

I/O Contract

Inputs

to_half:

Name	Type	Required	Description
tensor	torch.Tensor	Yes	The weight tensor to potentially convert
enable	bool	Yes	Whether to perform the FP16 conversion

read_metadata:

Name	Type	Required	Description
primary_model_name	str	Yes	Title/key of the primary model in checkpoints_list
secondary_model_name	str	Yes	Title/key of the secondary model (can be empty)
tertiary_model_name	str	Yes	Title/key of the tertiary model (can be empty)

Outputs

to_half:

Name	Type	Description
return	torch.Tensor	The tensor in FP16 if conversion was enabled and input was FP32; otherwise the original tensor unchanged

read_metadata:

Name	Type	Description
return	str	JSON string of merged metadata from all specified source models

Configuration Options

The following options are parameters to run_modelmerger that control output configuration:

Option	Type	Description
save_as_half	bool	Convert all float32 weight tensors to float16 before saving
checkpoint_format	str	Output file extension: "safetensors" or "ckpt"
bake_in_vae	str	Name of a VAE to embed into the merged model (looked up from sd_vae.vae_dict)
discard_weights	str	Regex pattern; any state dict key matching this pattern is removed from the output
save_metadata	bool	Whether to save metadata into the output file
add_merge_recipe	bool	Whether to include a machine-readable merge recipe in metadata
copy_metadata_fields	bool	Whether to copy metadata from source models
metadata_json	str	Additional custom JSON metadata to include

Usage Examples

Basic Usage

import torch
from modules.extras import to_half, read_metadata

# Convert a tensor to half precision
tensor = torch.randn(512, 512, dtype=torch.float32)
half_tensor = to_half(tensor, enable=True)
print(half_tensor.dtype)  # torch.float16

# No conversion when disabled
same_tensor = to_half(tensor, enable=False)
print(same_tensor.dtype)  # torch.float32

# Read merged metadata from source models
metadata_json = read_metadata(
    "v1-5-pruned-emaonly.safetensors [6ce0161689]",
    "dreamshaper_8.safetensors [abcd1234ef]",
    ""
)
print(metadata_json)  # JSON string of merged metadata

Related Pages

Implements Principle

• Principle:AUTOMATIC1111_Stable_diffusion_webui_Merge_output_options