LoRA

LoRA (Low-Rank Adaptation) adapts a pretrained model to a new task by freezing the original weights and training a small set of new parameters. Adaptation is faster and cheaper than full fine-tuning, which makes it practical for tasks like generating images in a new style.

LoRA checkpoints are typically only a couple hundred MBs. Load them with pipeline.load_lora_weights and set weight_name to the checkpoint filename. See load_lora_weights() for the shared parameters.

text-to-image

text-to-video

The pipeline LoRA APIs on this page work the same across model families. The denoising module is pipeline.unet on Stable Diffusion and SDXL, and pipeline.transformer on DiT-style models such as Flux, Stable Diffusion 3, AuraFlow, and LTX. The same pattern applies to video models such as CogVideoX, HunyuanVideo, and Wan. Model-level PEFT methods use that attribute and a matching prefix ("unet" or "transformer").

pipeline.load_lora_weights is the preferred way to load LoRA weights into the denoiser (UNet or transformer) and text encoders. See load_lora_weights() for the shared parameters. It can handle cases where:

the checkpoint shares keys across the denoiser and text encoders
the checkpoint uses separate denoiser and text-encoder keys

The load_lora_adapter() method loads a LoRA adapter at the model-level on any Diffusers model that subclasses PeftAdapterMixin, such as a UNet or transformer. It builds and prepares the adapter configuration and loads the weights into that module.

Use prefix to filter state dict keys for the target module ("unet" or "transformer"). The default is "transformer", so Stable Diffusion and SDXL must pass prefix="unet".

import torch
from diffusers import AutoPipelineForText2Image

pipeline = AutoPipelineForText2Image.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    dtype=torch.float16
).to("cuda")  # or "mps", "xpu", "cpu"
pipeline.unet.load_lora_adapter(
    "jbilcke-hf/sdxl-cinematic-1",
    weight_name="pytorch_lora_weights.safetensors",
    adapter_name="cinematic",
    prefix="unet"
)
# use cnmt in the prompt to trigger the LoRA
pipeline("A cute cnmt eating a slice of pizza, stunning color scheme, masterpiece, illustration").images[0]

import torch
from diffusers import FluxPipeline

pipeline = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev", dtype=torch.bfloat16
).to("cuda")  # or "mps", "xpu", "cpu"
pipeline.transformer.load_lora_adapter(
    "alvarobartt/ghibli-characters-flux-lora",
    adapter_name="lora",
    prefix="transformer",
)

torch.compile

torch.compile speeds up inference by compiling the PyTorch model to use optimized kernels. Before compiling, the LoRA weights need to be fused into the base model and unloaded first.

import torch
from diffusers import DiffusionPipeline

# load base model and LoRA
pipeline = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    dtype=torch.float16
).to("cuda")  # or "mps", "xpu", "cpu"
pipeline.load_lora_weights(
    "ostris/ikea-instructions-lora-sdxl",
    weight_name="ikea_instructions_xl_v1_5.safetensors",
    adapter_name="ikea"
)

# activate LoRA and set adapter weight
pipeline.set_adapters("ikea", adapter_weights=0.7)

# fuse LoRAs and unload weights
pipeline.fuse_lora(adapter_names=["ikea"], lora_scale=1.0)
pipeline.unload_lora_weights()

Typically, compile the denoiser because it’s the most compute intensive component of the pipeline.

unet

transformer

Refer to the hotswapping section to learn how to avoid recompilation when working with compiled models and multiple LoRAs.

Weight scale

Use the adapter_weights parameter on set_adapters() to control how much of a LoRA to apply. A weight of 0 is equivalent to only using the base model weights and a weight of 1 is equivalent to fully using the LoRA.

scalar weight

per-component scales

Scale scheduling

Dynamically adjusting the LoRA scale during sampling gives you better control over the overall composition and layout because certain steps may benefit more from an increased or reduced scale.

The character LoRA in the example below starts with a higher scale that gradually decays over the first 20 steps to establish the character generation. In the later steps, only a scale of 0.2 is applied to avoid adding too much of the LoRA features to other parts of the image the LoRA wasn’t trained on.

import torch
from diffusers import FluxPipeline

pipeline = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev", dtype=torch.bfloat16
).to("cuda")  # or "mps", "xpu", "cpu"

pipeline.load_lora_weights("alvarobartt/ghibli-characters-flux-lora", adapter_name="lora")

num_inference_steps = 30
lora_steps = 20
lora_scales = torch.linspace(1.5, 0.7, lora_steps).tolist()
lora_scales += [0.2] * (num_inference_steps - lora_steps + 1)

pipeline.set_adapters("lora", lora_scales[0])

def callback(pipeline: FluxPipeline, step: int, timestep: torch.LongTensor, callback_kwargs: dict):
    pipeline.set_adapters("lora", lora_scales[step + 1])
    return callback_kwargs

prompt = """
Ghibli style The Grinch, a mischievous green creature with a sly grin, peeking out from behind a snow-covered tree while plotting his antics, 
in a quaint snowy village decorated for the holidays, warm light glowing from cozy homes, with playful snowflakes dancing in the air
"""
pipeline(
    prompt=prompt,
    guidance_scale=3.0,
    num_inference_steps=num_inference_steps,
    generator=torch.Generator().manual_seed(42),
    callback_on_step_end=callback,
).images[0]

Hotswapping

Hotswapping LoRAs is an efficient way to work with multiple LoRAs while avoiding accumulating memory from multiple calls to load_lora_weights and in some cases, recompilation, if a model is compiled. This workflow requires a loaded LoRA because the new LoRA weights are swapped in place for the existing loaded LoRA.

import torch
from diffusers import DiffusionPipeline

# load base model and LoRAs
pipeline = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    dtype=torch.float16
).to("cuda")  # or "mps", "xpu", "cpu"
pipeline.load_lora_weights(
    "ostris/ikea-instructions-lora-sdxl",
    weight_name="ikea_instructions_xl_v1_5.safetensors",
    adapter_name="ikea"
)

Hotswapping is unsupported for LoRAs that target the text encoder.

Set hotswap=True on load_lora_weights to swap the second LoRA. Use the adapter_name parameter to indicate which LoRA to swap. If you omit adapter_name when loading, Diffusers names adapters default_0, default_1, and so on. Named adapters are easier to target explicitly.

pipeline.load_lora_weights(
    "lordjia/by-feng-zikai",
    hotswap=True,
    adapter_name="ikea"
)

Compiled models

For compiled models, use enable_lora_hotswap() to avoid recompilation when hotswapping LoRAs. This method should be called before loading the first LoRA and torch.compile should be called after loading the first LoRA.

The enable_lora_hotswap() method isn’t always necessary if the second LoRA targets the identical LoRA ranks and scales as the first LoRA.

Within enable_lora_hotswap(), set target_rank to the highest rank among the adapters you plan to load (default 128).

import torch
from diffusers import DiffusionPipeline

# load base model and LoRAs
pipeline = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    dtype=torch.float16
).to("cuda")  # or "mps", "xpu", "cpu"
# 1. enable_lora_hotswap
pipeline.enable_lora_hotswap(target_rank=128)
pipeline.load_lora_weights(
    "ostris/ikea-instructions-lora-sdxl",
    weight_name="ikea_instructions_xl_v1_5.safetensors",
    adapter_name="ikea"
)
# 2. torch.compile
pipeline.unet = torch.compile(pipeline.unet, mode="reduce-overhead", fullgraph=True)

# 3. hotswap
pipeline.load_lora_weights(
    "lordjia/by-feng-zikai",
    hotswap=True,
    adapter_name="ikea"
)

Move your code inside the with torch._dynamo.config.patch(error_on_recompile=True) context manager to detect if a model was recompiled. If a model is recompiled despite following all the steps above, please open an issue with a reproducible example.

If you expect varied resolutions during inference with this feature, then make sure to set dynamic=True during compilation. Refer to dynamic shape compilation for more details.

There are still scenarios where recompilation is unavoidable, such as when the hotswapped LoRA targets more layers than the initial adapter. Try to load the LoRA that targets the most layers first. For more details about this limitation, refer to the PEFT hotswapping docs.

Why `target_rank` and load order matter

The enable_lora_hotswap() method converts the LoRA scaling factor from floats to torch.tensors and pads the shape of the weights to the largest required shape to avoid reassigning the whole attribute when the data in the weights are replaced.

This is why the target_rank argument is important. The results are unchanged even when the values are padded with zeros. Computation may be slower though depending on the padding size.

Since no new LoRA attributes are added, each subsequent LoRA is only allowed to target the same layers, or subset of layers, the first LoRA targets. Choosing the LoRA loading order is important because if the LoRAs target disjoint layers, you may end up creating a dummy LoRA that targets the union of all target layers.

See the LoRA fast blog post for more on LoRA hotswapping with compiled models, and for more implementation details, take a look at the hotswap.py file.

Merge

The weights from each LoRA can be merged together to produce a blend of multiple existing styles. There are several methods for merging LoRAs, each of which differ in how the weights are merged (may affect generation quality).

set_adapters

The set_adapters() method merges LoRAs by concatenating their weighted matrices. Pass the LoRA names to set_adapters() and use the adapter_weights parameter to control the scaling of each LoRA. For example, if adapter_weights=[0.5, 0.5], the output is an average of both LoRAs.

Prefer adapter_weights on set_adapters() to control how much of each merged LoRA to apply. See the Weight scale section for more details.

import torch
from diffusers import DiffusionPipeline

pipeline = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    dtype=torch.float16
).to("cuda")  # or "mps", "xpu", "cpu"
pipeline.load_lora_weights(
    "ostris/ikea-instructions-lora-sdxl",
    weight_name="ikea_instructions_xl_v1_5.safetensors",
    adapter_name="ikea"
)
pipeline.load_lora_weights(
    "lordjia/by-feng-zikai",
    weight_name="fengzikai_v1.0_XL.safetensors",
    adapter_name="feng"
)
pipeline.set_adapters(["ikea", "feng"], adapter_weights=[0.7, 0.8])
# use by Feng Zikai to activate the lordjia/by-feng-zikai LoRA
pipeline("A bowl of ramen shaped like a cute kawaii bear, by Feng Zikai").images[0]

add_weighted_adapter

This is an experimental method and you can refer to PEFT’s Model merging for more details. Take a look at this issue if you’re interested in the motivation and design behind this integration.

The add_weighted_adapter method enables more efficient merging methods like TIES or DARE. These merging methods remove redundant and potentially interfering parameters from merged models. Keep in mind the LoRA ranks need to have identical ranks to be merged. The walkthrough below uses SDXL’s UNet.

Make sure the latest stable version of Diffusers and PEFT is installed.

pip install -U -q diffusers peft

Load a UNet that corresponds to the LoRA UNet.

import copy
import torch
from diffusers import AutoModel, DiffusionPipeline
from peft import get_peft_model, LoraConfig, PeftModel

unet = AutoModel.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    dtype=torch.float16,
    use_safetensors=True,
    variant="fp16",
    subfolder="unet",
).to("cuda")  # or "mps", "xpu", "cpu"

Load a pipeline, pass the UNet to it, and load a LoRA.

pipeline = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    variant="fp16",
    dtype=torch.float16,
    unet=unet
).to("cuda")  # or "mps", "xpu", "cpu"
pipeline.load_lora_weights(
    "ostris/ikea-instructions-lora-sdxl",
    weight_name="ikea_instructions_xl_v1_5.safetensors",
    adapter_name="ikea"
)

Create a PeftModel from the LoRA checkpoint by combining the first UNet you loaded and the LoRA UNet from the pipeline.

sdxl_unet = copy.deepcopy(unet)
ikea_peft_model = get_peft_model(
    sdxl_unet,
    pipeline.unet.peft_config["ikea"],
    adapter_name="ikea"
)

original_state_dict = {f"base_model.model.{k}": v for k, v in pipeline.unet.state_dict().items()}
ikea_peft_model.load_state_dict(original_state_dict, strict=True)

You can save and reuse the ikea_peft_model by pushing it to the Hub as shown below.
ikea_peft_model.push_to_hub("ikea_peft_model", token=TOKEN)

Repeat this process and create a PeftModel for the second LoRA.

pipeline.delete_adapters("ikea")
sdxl_unet.delete_adapters("ikea")

pipeline.load_lora_weights(
    "lordjia/by-feng-zikai",
    weight_name="fengzikai_v1.0_XL.safetensors",
    adapter_name="feng"
)
pipeline.set_adapters(adapter_names="feng")

feng_peft_model = get_peft_model(
    sdxl_unet,
    pipeline.unet.peft_config["feng"],
    adapter_name="feng"
)

original_state_dict = {f"base_model.model.{k}": v for k, v in pipeline.unet.state_dict().items()}
feng_peft_model.load_state_dict(original_state_dict, strict=True)

Load a base UNet model and load the adapters.

base_unet = AutoModel.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    dtype=torch.float16,
    use_safetensors=True,
    variant="fp16",
    subfolder="unet",
).to("cuda")  # or "mps", "xpu", "cpu"

model = PeftModel.from_pretrained(
    base_unet,
    "stevhliu/ikea_peft_model",
    use_safetensors=True,
    subfolder="ikea",
    adapter_name="ikea"
)
model.load_adapter(
    "stevhliu/feng_peft_model",
    use_safetensors=True,
    subfolder="feng",
    adapter_name="feng"
)

Merge the LoRAs with add_weighted_adapter and specify how you want to merge them with combination_type. The example below uses the "dare_linear" method (refer to this blog post to learn more about these merging methods), which randomly prunes some weights and then performs a weighted sum of the tensors based on the set weightage of each LoRA in weights.

Activate the merged LoRAs with set_adapters().

model.add_weighted_adapter(
    adapters=["ikea", "feng"],
    combination_type="dare_linear",
    weights=[1.0, 1.0],
    adapter_name="ikea-feng"
)
model.set_adapters("ikea-feng")

pipeline = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    unet=model,
    variant="fp16",
    dtype=torch.float16,
).to("cuda")  # or "mps", "xpu", "cpu"
pipeline("A bowl of ramen shaped like a cute kawaii bear, by Feng Zikai").images[0]

fuse_lora

The fuse_lora() method fuses the LoRA weights directly with the original denoiser (UNet or transformer) and text encoder weights of the underlying model. This reduces the overhead of loading the underlying model for each LoRA because it only loads the model once, which lowers memory usage and increases inference speed.

import torch
from diffusers import DiffusionPipeline

pipeline = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    dtype=torch.float16
).to("cuda")  # or "mps", "xpu", "cpu"
pipeline.load_lora_weights(
    "ostris/ikea-instructions-lora-sdxl",
    weight_name="ikea_instructions_xl_v1_5.safetensors",
    adapter_name="ikea"
)
pipeline.load_lora_weights(
    "lordjia/by-feng-zikai",
    weight_name="fengzikai_v1.0_XL.safetensors",
    adapter_name="feng"
)
pipeline.set_adapters(["ikea", "feng"], adapter_weights=[0.7, 0.8])

Call fuse_lora() to fuse them. The lora_scale parameter controls how much to scale the output by with the LoRA weights. It is important to make this adjustment now because passing scale to cross_attention_kwargs won’t work in the pipeline.

pipeline.fuse_lora(adapter_names=["ikea", "feng"], lora_scale=1.0)

Unload the LoRA weights since they’re already fused with the underlying model. Save the fused pipeline with either save_pretrained() to save it locally or ~PushToHubMixin.push_to_hub to save it to the Hub.

save locally

save to Hub

The fused pipeline can now be quickly loaded for inference without requiring each LoRA to be separately loaded.

pipeline = DiffusionPipeline.from_pretrained(
    "username/fused-ikea-feng", dtype=torch.float16,
).to("cuda")  # or "mps", "xpu", "cpu"
pipeline("A bowl of ramen shaped like a cute kawaii bear, by Feng Zikai").images[0]

Use unfuse_lora() to restore the underlying model’s weights, for example, if you want to use a different lora_scale value. You can only unfuse if there is a single LoRA fused. For example, it won’t work with the pipeline from above because there are multiple fused LoRAs. In these cases, you’ll need to reload the entire model.

pipeline.unfuse_lora()

Manage

Diffusers provides several methods to help you manage working with LoRAs. These methods can be especially useful if you’re working with multiple LoRAs.

set_adapters

set_adapters() also activates the current LoRA to use if there are multiple active LoRAs. This allows you to switch between different LoRAs by specifying their name.

import torch
from diffusers import DiffusionPipeline

pipeline = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    dtype=torch.float16
).to("cuda")  # or "mps", "xpu", "cpu"
pipeline.load_lora_weights(
    "ostris/ikea-instructions-lora-sdxl",
    weight_name="ikea_instructions_xl_v1_5.safetensors",
    adapter_name="ikea"
)
pipeline.load_lora_weights(
    "lordjia/by-feng-zikai",
    weight_name="fengzikai_v1.0_XL.safetensors",
    adapter_name="feng"
)
# activates the feng LoRA instead of the ikea LoRA
pipeline.set_adapters("feng")

save_lora_adapter

Save an adapter with save_lora_adapter() on the same module you loaded into (pipeline.unet or pipeline.transformer).

import torch
from diffusers import AutoPipelineForText2Image

pipeline = AutoPipelineForText2Image.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    dtype=torch.float16
).to("cuda")  # or "mps", "xpu", "cpu"
pipeline.unet.load_lora_adapter(
    "jbilcke-hf/sdxl-cinematic-1",
    weight_name="pytorch_lora_weights.safetensors",
    adapter_name="cinematic",
    prefix="unet"
)
pipeline.unet.save_lora_adapter("path/to/save", adapter_name="cinematic")

unload_lora_weights

The unload_lora_weights() method unloads any LoRA weights in the pipeline to restore the underlying model weights.

pipeline.unload_lora_weights()

disable_lora

The disable_lora() method disables all LoRAs (but they’re still kept on the pipeline) and restores the pipeline to the underlying model weights.

pipeline.disable_lora()

get_active_adapters

The get_active_adapters() method returns a list of active LoRAs attached to a pipeline.

pipeline.get_active_adapters()
# ["cereal", "ikea"]

get_list_adapters

The get_list_adapters() method returns all loaded LoRAs for each component in the pipeline. The dictionary keys match the loaded modules, such as unet or transformer and any text encoders.

pipeline.get_list_adapters()
# {"unet": ["cereal", "ikea"], "text_encoder_2": ["cereal"]}

delete_adapters

The delete_adapters() method completely removes a LoRA and its layers from a model.

pipeline.delete_adapters("ikea")

Next steps

Browse the LoRA Studio for different LoRAs to use or you can upload your favorite LoRAs from Civitai to the Hub with the Space below.

You can find additional LoRAs in the FLUX LoRA the Explorer and LoRA the Explorer Spaces.
Check out the Fast LoRA inference for Flux with Diffusers and PEFT blog post to learn how to optimize LoRA inference with methods like FlashAttention-3 and fp8 quantization.

Update on GitHub