Models, LoRA & Extensions

Can You Combine Multiple LoRAs in One Stable Diffusion Image?

PhoneDiffusion
Can You Combine Multiple LoRAs in One Stable Diffusion Image?

Models & LoRA

Yes. Stable Diffusion can use multiple LoRAs in one image: add one LoRA tag per adapter in AUTOMATIC1111, chain LoraLoader nodes in ComfyUI, or activate several named adapters with their own weights in Diffusers. Every LoRA must match the base-model family, and lower weights usually combine better than full strength.

The short answer

Yes, you can combine multiple LoRAs in one Stable Diffusion image. Each LoRA is a small patch applied on top of the base model, and the common tools let you stack several patches and give each its own strength.

The catch is compatibility. Every LoRA must be trained for the same base-model family as the checkpoint you generate with: an SDXL LoRA does not belong on an SD 1.5 checkpoint. And two LoRAs that push on the same features (face, style, lighting) can fight each other, so stacking works best at reduced weights.

How to use multiple LoRAs in AUTOMATIC1111 and Forge

The AUTOMATIC1111 web UI activates a LoRA with a prompt tag of the form <lora:filename:multiplier>, placed anywhere in the prompt. The wiki describes the multiplier as a number, generally from 0 to 1, that sets how strongly the LoRA affects the output. To combine LoRAs, add one tag per LoRA, for example a style LoRA and a character LoRA each with its own multiplier.

Keep any trigger words each LoRA's author documents in the prompt as well. The wiki also notes a batch limitation: a batch with multiple different prompts only uses the LoRA from the first prompt, so test combinations with one prompt per batch.

How to chain multiple LoRAs in ComfyUI

ComfyUI loads LoRA files from the models/loras folder with a LoraLoader node, which patches both the MODEL and the CLIP model. The official LoRA example applies multiple LoRAs by chaining LoraLoader nodes: the MODEL and CLIP outputs of the first loader feed the second loader, and the last loader feeds the sampler and the prompt encoders.

Each loader has its own strength, so you can lower one LoRA without touching the other. Because the chain is part of the workflow, it is saved with the workflow JSON. Anyone rerunning it needs the same checkpoint and LoRA files with the same names, because the workflow does not include model files.

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

How to combine LoRA adapters in Diffusers

In Hugging Face Diffusers, load each LoRA with load_lora_weights and give it an adapter_name, then activate them together with set_adapters, passing the names and an adapter_weights list. The documentation's SDXL example combines two adapters with set_adapters(["ikea", "feng"], adapter_weights=[0.7, 0.8]), and notes that weights of [0.5, 0.5] average the two LoRAs.

get_active_adapters shows which LoRAs are on, disable_lora turns them all off without unloading them, and delete_adapters removes one completely. One limit to know: set_adapters only scales attention weights; other LoRA layers such as ResNets keep a scale of 1.0.

Stacking, fusing, and merging are different operations

Stacking keeps each LoRA separate and adds their influence at generation time, so you can reweight or remove one at any point. Diffusers' fuse_lora instead bakes the active LoRAs into the base model's weights, which lowers memory use and speeds up inference, but unfuse_lora only works when a single LoRA was fused; with several fused LoRAs you reload the model.

Merging creates a new adapter from several. Diffusers documents an experimental add_weighted_adapter path with methods such as TIES and DARE that prune redundant or interfering parameters; it requires the LoRAs to have identical ranks. A UI button called ‘merge’ may do something different, so check what your tool actually does; baking a LoRA into a full model is a separate job, covered in checkpoint merge vs LoRA.

A controlled way to find weights that work together

Generate a baseline with the base model and a fixed prompt, seed, sampler, steps, and size. Test LoRA A alone, then LoRA B alone, each below full strength. Only then enable both without changing anything else.

If the combined image loses the subject, style, or anatomy you wanted, lower one weight at a time instead of adding more prompt words; Stable Diffusion LoRA weights explained covers what each weight number does. The best pair is rarely both LoRAs at full strength. Record the base model, the exact LoRA files and versions, trigger words, weights, prompt, and seed with every result you keep.

PhoneDiffusion availability and article scope

PhoneDiffusion uses curated, converted Core ML model packs from its in-app catalog for on-device generation, and its optional cloud models (paid monthly or yearly plans) come from a fixed in-app list. It does not provide a general production importer for arbitrary LoRA files, checkpoints, ControlNet files, or models downloaded from Civitai or Hugging Face, and the cloud models take a text prompt only.

This guide explains the wider Stable Diffusion ecosystem. Use a desktop or server tool with documented multi-LoRA support when combining LoRAs is essential; do not assume that a desktop LoRA can be installed in the current iPhone app.

Field notes

  • Multiple LoRAs work when every LoRA matches the checkpoint's base-model family.
  • AUTOMATIC1111 uses one <lora:filename:multiplier> tag per LoRA; ComfyUI chains LoraLoader nodes; Diffusers uses set_adapters with adapter_weights.
  • Test each LoRA alone first, then combine at reduced weights with the same prompt and seed.
  • Stacking keeps LoRAs adjustable; fusing and merging bake them in.

Try it on your phone

  1. Confirm that every LoRA targets the same base-model family as the selected checkpoint.
  2. Save a base-model result with a fixed prompt, seed, sampler, steps, and size.
  3. Test each LoRA alone below full strength, with its trigger words.
  4. Enable all of them together (one tag per LoRA, chained LoraLoader nodes, or set_adapters with a weight list).
  5. Lower one weight at a time until the features stop competing, and save the exact LoRA versions and weights with the result.

Sources and verification

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