Models, LoRA & Extensions

Checkpoint Merge vs LoRA: Better Ways to Customize Stable Diffusion

PhoneDiffusion
Checkpoint Merge vs LoRA: Better Ways to Customize Stable Diffusion

Models & LoRA

A checkpoint merge blends two or more full Stable Diffusion models into one new model file, while a LoRA is a small add-on you load on top of a base model at a weight you choose. Use a LoRA when you want a style or character you can turn up, down or off; merge a LoRA into a checkpoint only when you want that change permanent in a single file.

Checkpoint merge vs LoRA: the short answer

Both change what a Stable Diffusion model draws, but in different places. A checkpoint merge does arithmetic on the full weights of whole models and saves the result as a new checkpoint. A LoRA leaves the base model untouched: the original LoRA paper describes freezing the pretrained weights and training small low-rank matrices that are added to them.

That makes a LoRA the flexible choice. Hugging Face's Diffusers documentation notes that LoRA files are typically a couple hundred MB, far smaller than a full model, and that a weight of 0 means only the base model while 1 means the full LoRA. A merge is the permanent choice: once the weights are combined, the result is simply a new model.

What a checkpoint merge does

In the AUTOMATIC1111 web UI, the Checkpoint Merger tab offers two main methods. Weighted sum blends model A and model B as (1 − M) × A + M × B, so a multiplier M of 0.3 keeps 70% of A. Add difference computes A + M × (B − C): it takes what model B learned on top of model C and adds it to model A, which is why it needs a third (tertiary) model.

ComfyUI does the same with nodes under advanced > model_merging. Its examples show a simple two-model merge, merges where the input, middle and output blocks of the UNet each get their own ratio, and the add-difference recipe (inpaint model − base model) × 1.0 + other model for turning a custom checkpoint into an inpainting one. A CheckpointSave node writes the result to output/checkpoints.

What a LoRA does

A LoRA is a patch for one model family. You load it with a base model and set its strength, and you can stack several: Diffusers' set_adapters takes a list of LoRAs and a weight for each, so you can rebalance or drop one at any time without touching the base model.

The trade-off is that each LoRA is an extra file the app has to load and apply. If you always use the same base model with the same LoRAs at the same weights, baking them in can make sense.

PhoneDiffusion generates on-device on iPhone or iPad, with optional cloud models. Download on the App Store →

How to merge a LoRA into a checkpoint

In Diffusers, load the LoRA, then call fuse_lora with the lora_scale you want; the documentation says to set the scale at this point, because a scale passed later at generation time no longer applies. Fusing lowers memory use and speeds up inference. Then call unload_lora_weights and save_pretrained to write the fused model. unfuse_lora can undo the fuse only when a single LoRA was fused.

In ComfyUI, the model merging examples note that LoRAs are a patch on the model weights, so they can be merged into the model: load the checkpoint and the LoRA, then send the patched model to CheckpointSave. On the command line, kohya-ss sd-scripts has networks/merge_lora.py for SD 1.x and 2.x (add --v2 for 2.x) and networks/sdxl_merge_lora.py for SDXL. Both take the checkpoint (--sd_model), one or more LoRAs (--models) with a ratio for each (--ratios), and an output file (--save_to). Leave out --sd_model and merge_lora.py merges the LoRAs into one LoRA instead.

Whichever tool you use, test the LoRA weight first and keep the original files. After baking, the LoRA's strength is fixed; to change it, you go back to the base model and merge again.

Which should you use?

Use a LoRA when you switch styles, characters or concepts often, want to fine-tune the strength per image, or are short on storage. Use a checkpoint merge when you want a new all-round base model (for example a blend of a realistic and an illustrated model), or when a tool or workflow needs one model file with no separate LoRA.

Two cautions. Merges combine weights layer by layer, so blend only models of the same family (two SD 1.5 models, or two SDXL models), and use a LoRA with the model family it was made for. And check the license of every model and LoRA in a merge before you share the result.

PhoneDiffusion availability and article scope

Checkpoint merging and LoRA fusing are desktop workflows. PhoneDiffusion does not provide a general production importer for custom checkpoints, merged models or LoRA files; it runs curated Core ML model packs (SD 1.5, SD 2.1, and SDXL) on supported iPhone and iPad devices with iOS 17 or later, where local generation is private and works offline once a pack is downloaded.

It also offers optional cloud models, such as FLUX and Seedream, on paid monthly or yearly plans; those are text-to-image only, and the text prompt is sent online to generate the image.

Field notes

  • A checkpoint merge combines whole models into a new model file.
  • A LoRA is a small add-on with an adjustable weight; 0 means the base model only.
  • Weighted sum blends two models; Add difference needs a third model.
  • Merging a LoRA into a checkpoint fixes its strength, so test the weight first.
  • Only merge models and LoRAs from the same model family.

Try it on your phone

  1. Decide whether you need flexibility (LoRA) or one fixed model file (merge).
  2. Test the LoRA at a few weights with a fixed prompt and seed.
  3. Merge with your tool: fuse_lora in Diffusers, CheckpointSave in ComfyUI, or merge_lora.py (sdxl_merge_lora.py for SDXL).
  4. Save the merged model as a new file and keep the originals.
  5. Compare the merged model with the base model plus LoRA on the same seed.

Sources and verification

Primary and maintainer documentation checked for this guide. Product-specific details were reviewed against the current PhoneDiffusion scope.