To use a LoRA in Fooocus, place an SDXL-compatible adapter in models/loras, refresh the file list, and select and enable it under Advanced → Models. Fooocus provides a weight control for each configured LoRA slot. This loads an already trained adapter for generation; creating adapter weights is a separate training workflow. config training
What are the key facts about Fooocus LoRA loading?
| Field | Verified detail |
|---|---|
| Developer | Fooocus application: lllyasviel and project contributors; each external LoRA has its own author. readme |
| Released | No separate Fooocus LoRA product release is documented; LoRA loading is an application capability recorded in the official source and update log. ui |
| Type | Inference-time use of an already trained adapter within an image-generation application. ui |
| Size or parameters | No universal adapter size published; LoRA configuration, including rank, determines the trainable parameter set. training |
| License and access | Fooocus code: GPL-3.0; inspect the chosen checkpoint and LoRA's own distribution terms. readme |
| Where it runs | Inside the Fooocus process; adapter files are loaded from the configured LoRA directory, defaulting to models/loras. config
|
How does Fooocus apply LoRAs during image generation?
Fooocus makes adapter selection explicit in its Models panel. The source defines a file dropdown, an enable checkbox, and a Weight slider for each configured LoRA slot. ui
That arrangement supports a straightforward comparison. Generate a baseline with the adapter disabled, enable it, and inspect whether the result moves toward the appearance you intended.
The application also exposes configuration for LoRA directories and default adapter selections. Those settings are useful when you want to retain a known setup between sessions. config
LoRA training and LoRA loading serve different purposes. Hugging Face's training guide describes optimizing adapter weights and then loading the resulting artifact into a generation pipeline. training
In Fooocus, your immediate task is to test the supplied adapter. You do not need to assemble a dataset or choose training epochs simply to compare an already trained LoRA with an unmodified baseline.
The Fooocus setup guide provides installation context. Keep this adapter test small until the ordinary image-generation workflow already works on your machine.
Which LoRAs are compatible with Fooocus?
The Fooocus interface labels its base-model selector as SDXL-only, while the refiner selector has a separate SDXL-or-SD-1.5 label. That distinction should guide your first compatibility check. ui
For an SDXL base-model trial, choose an adapter documented for that architecture. A filename containing “LoRA” or ending in .safetensors does not by itself identify which model the adapter was trained to modify. training
Read the adapter author's model card before experimenting. Check the intended base, usage instructions, license, and any required prompt terms. If those details are absent, record the uncertainty rather than inventing a recommended setting.
The application offers weight controls without publishing one ideal value for every external adapter. Use the selected adapter's own instructions as the initial reference and evaluate changes against your baseline. ui
Treat several enabled adapters as a combined experiment. If the result becomes difficult to explain, return to one adapter and reintroduce the others individually.
Fooocus's Enhance documentation also describes interactions between inpaint-engine LoRAs and performance LoRAs. Check those documented constraints before combining accelerated generation with a refinement workflow. enhance
The current README describes limited long-term support focused on bug fixes. Verify that an online example applies to the official application version you use, especially if it demonstrates a fork or a different model architecture. readme
How do you load and configure a LoRA in Fooocus?
Begin with a working Fooocus installation and a downloaded SDXL-compatible adapter. Keep the adapter's original filename and source information for this first test.
- Place the adapter file in
Fooocus/models/loras, or the custom location configured throughpath_loras. The configuration source identifies that default directory. config - Launch Fooocus using the official launcher or the documented Python entry point from its active environment. readme
- Open Advanced, then Models, and select the intended SDXL base checkpoint. Use Refresh All Files if the newly added adapter is missing from the dropdown. ui
- Select the adapter in a LoRA slot, enable it, and set its weight according to the adapter author's documented starting point. ui
- Generate a baseline with the adapter disabled, then an adapter-enabled comparison. Reuse the same prompt and seed, and inspect the difference against the effect you wanted.
For a Python installation whose documented environment is already active, the launch command is:
python entry_with_update.py
The configuration source also supports persistent defaults.
This illustrative JSON entry uses a placeholder filename and example weight; merge it into your existing configuration only after substituting your real adapter and chosen value. config
{
"default_loras": [[true, "your-sdxl-adapter.safetensors", 0.5]]
}
The example weight is not a universal recommendation. First establish a useful result through the interface, then record that tested value as your default.
Use a simple diagnostic sequence when something fails. Start by confirming whether the file appears in the list, whether it is enabled, and whether generation reports a loading error. These are separate observations.
If the file is absent, recheck the active LoRA directory and refresh the file list.
The official configuration permits custom paths, so checking a different installation's folder will not establish where the running application looks. config
If loading fails, return to the adapter's documented base-model requirements. Keep the error message and identify the exact adapter and checkpoint before replacing files or changing generation settings.
If loading succeeds but the effect is unclear, use a prompt that makes the intended style or subject visible. Disable other adapters for the comparison and change only the selected adapter's weight between trials.
For reference-image character work, see the separate Fooocus virtual-model guide.
That workflow gives you another way to test a portrait reference without making adapter training part of the task.
How do Fooocus LoRA loading and LoRA training differ?
| Route | Main task |
|---|---|
| Fooocus LoRA loading | Select existing adapters and their weights through the Models panel. ui |
| ComfyUI LoRA workflow | Arrange adapter loading within a node-based generation workflow. comfy |
| Diffusers LoRA training | Prepare a training environment and optimize new adapter weights before inference. training |
Use the ComfyUI LoRA guide for the graph-based route. Choose a training workflow only when you need to create an adapter rather than apply one.
What else should you know about Fooocus LoRAs?
Where do I put LoRA files in Fooocus?
Fooocus uses models/loras by default. If you configured path_loras, use that location, then click Refresh All Files in the Models panel. config
Does loading a LoRA train Fooocus on my images?
Fooocus's Models panel selects an existing adapter and its weight for generation. Training a LoRA requires a separate procedure that optimizes and saves adapter weights. ui
What LoRA weight should I use in Fooocus?
Fooocus provides a Weight slider for each configured adapter slot. Start from the adapter author's instructions and compare with that adapter disabled before deciding which setting suits your image. ui
Can I enable several LoRAs in Fooocus?
Fooocus supports multiple configured LoRA slots, each with its own enable control and weight. Test adapters individually before comparing a combined setup. ui
Sources
- Fooocus official README
- Official Models panel implementation
- Official LoRA path and defaults configuration
- Fooocus update log
- Diffusers LoRA training documentation
- Maintainer's Enhance compatibility notes
- ComfyUI official repository
Top comments (0)