Yes, compatible Stable Diffusion pipelines can activate or merge multiple LoRA adapters and assign each a weight. Every adapter must match the base-model family and supported layers; combinations can conflict and use more memory.
The short answer
Yes. A pipeline that supports multiple adapters can load more than one LoRA, give each adapter a name, activate them together, and set separate weights. Hugging Face Diffusers documents this with multiple adapter names and weights through its adapter APIs.
That does not mean any two LoRA files will work together. Each must be compatible with the same base-model family and the pipeline components it targets. A LoRA trained for SDXL is not automatically compatible with an SD 1.5 checkpoint, and a tool may support LoRA on one component but not another.
Active adapters and merged weights are not identical workflows
Some pipelines keep adapters separate and combine their influence during inference. Others fuse or merge adapter weights into the base model or into a new adapter. Separate adapters are easier to turn on, turn off, and reweight; fused weights may simplify a repeated deployment but are harder to undo and need careful saving.
Use the method documented by the runtime you actually run. A UI label such as ‘merge’ may describe a different operation from a library method with the same everyday name.
A controlled way to combine two LoRAs
First generate a baseline with the base model and a fixed prompt and seed. Activate LoRA A at a conservative weight and save the result. Return to the baseline, test LoRA B alone, then activate both without changing the prompt, seed, scheduler, or size.
If the combined result loses the subject, style, or anatomy you wanted, reduce one weight rather than adding more prompt tokens. Two adapters can compete for the same visual features. The best pair is not necessarily the pair with the highest weights.
Memory and reproducibility limits
Loading multiple adapters adds memory pressure. Some runtimes can offload inactive adapters or hotswap one adapter for another; those techniques are designed for switching, not necessarily for using every adapter simultaneously. Compiled pipelines can also have adapter-specific limitations.
Record the exact base model, adapter files and versions, trigger phrases, weights, prompt, seed, and runtime. A recipe that omits the base model and adapter versions is not reproducible enough to debug.
PhoneDiffusion availability and article scope
PhoneDiffusion uses curated, converted Core ML model packs from its in-app catalog. It does not provide a general production importer for arbitrary LoRA files, checkpoints, ControlNet files, or models downloaded from Civitai or Hugging Face.
This guide explains the wider Stable Diffusion ecosystem. Use a desktop or server pipeline with documented multi-adapter support when combining LoRAs is essential; do not assume that a desktop adapter can be installed in the current iPhone app.
Field notes
- Multiple LoRAs work only when the base model, targeted components, and runtime are compatible.
- Test each adapter alone before combining them, using the same prompt and seed.
- Adapter weights are controls, not quality scores; high values can conflict.
Try it on your phone
- Confirm that both LoRAs target the same base-model family as the selected checkpoint.
- Save a base-model result with a fixed prompt, seed, scheduler, and size.
- Test each LoRA alone at a conservative weight.
- Activate both, change one weight at a time, and save the exact adapter versions with the result.
Sources and verification
Primary and maintainer documentation checked for this guide. Product-specific details were reviewed against the current PhoneDiffusion scope.
- Hugging Face Diffusers — loading and combining LoRA adaptersMaintainer documentation for adapter compatibility, names, weights, multiple adapters, hotswapping, and memory tradeoffs.