Using Hugging Face LoRAs

Hugging Face is the largest open-source model hub and hosts thousands of LoRAs for Flux, SDXL, and SD 1.5. Reference them by repo ID (e.g., alvdansen/frosting_lane_flux) in the LoRA URL field. LoRA weights are stored as .safetensors files, and most repos include trigger words and example images in their model card.

Finding LoRAs on Hugging Face

Hugging Face hosts the largest collection of open-source LoRA weights. The best starting point is the filtered model listing at huggingface.co/models?pipeline_tag=text-to-image&other=lora, which shows only text-to-image LoRA adapters.

Filter by tags: add flux, sdxl, or stable-diffusion to narrow results to a specific architecture.
Sort by downloads, trending, or likes to find the most popular and battle-tested models.
For interactive discovery, the "FLUX LoRA the Explorer" Space at huggingface.co/spaces/multimodalart/flux-lora-the-explorer lets you preview and compare Flux LoRAs side-by-side before committing to one.
Each model page includes a model card (README) with usage instructions, trigger words, and example outputs generated by the creator.

Repository Structure

A typical Hugging Face LoRA repository contains a small set of files:

`pytorch_lora_weights.safetensors` or `adapter_model.safetensors`

The actual LoRA weight file. Typically between 2 MB and 200 MB depending on the rank and target modules. This is the file that gets loaded at inference time.

`adapter_config.json`

LoRA configuration metadata including rank, alpha, target modules, and the base model reference. This tells the loader how to apply the weights correctly.

`README.md`

The model card with trigger words, base model info, training details, and example images. Always read this before using a LoRA.

The standard file format is .safetensors — a safe, fast serialization format that prevents arbitrary code execution during loading. It is the default for all modern LoRAs. Legacy .bin or .ckpt files still exist in older repositories, but safetensors takes priority when both are present.

How to Reference HF LoRAs

In CSF, paste the Hugging Face repo ID into the LoRA URL field in the Integrations modal. The format is simply owner/model-name. For example:

alvdansen/frosting_lane_flux

This works because Replicate's hf_lora parameter natively resolves Hugging Face repo IDs. When Replicate sees a repo ID, it fetches the safetensors file directly from the Hugging Face hub, caches it, and applies it to the base model at inference time.

For repositories that contain multiple LoRA files or use subfolder structures, you may need to verify that the default file path resolves correctly. Most well-maintained repos follow the standard naming convention and work out of the box.

Checking Base Model Compatibility

LoRAs only work with the exact base model architecture they were trained on. A Flux LoRA applied to an SDXL model will produce broken, incoherent results — or fail to load entirely. This is the single most common mistake when using LoRAs.

To verify compatibility, check the model card's base_model field. Common values you will see:

Flux.1 Dev

black-forest-labs/FLUX.1-dev — the most common base for new Flux LoRAs. High quality, supports guidance.

Flux.1 Schnell

black-forest-labs/FLUX.1-schnell — faster distilled variant. Fewer LoRAs target this base.

SDXL 1.0

stabilityai/stable-diffusion-xl-base-1.0 — large ecosystem of existing LoRAs from the SDXL era.

SD 1.5

stable-diffusion-v1-5/stable-diffusion-v1-5 — legacy but still has the largest total LoRA count.

If the model card does not explicitly specify a base model, check the repository tags or the description for mentions of Flux, SDXL, or SD 1.5.

Popular Hugging Face LoRAs

Here are some well-known, community-tested LoRAs available on Hugging Face:

`XLabs-AI/flux-RealismLora`

Photorealism enhancement for Flux. Improves skin texture, lighting, and overall photographic quality. No trigger word needed — activates automatically.

`alvdansen/frosting_lane_flux`

Stylized frosted/pastel aesthetic for Flux. Produces soft, dreamy imagery with muted color palettes and gentle lighting.

`Shakker-Labs/FLUX.1-dev-LoRA-Dark-Fantasy`

Dark fantasy art style with dramatic shadows, gothic environments, and moody atmospheric lighting.

`black-forest-labs/FLUX.1-Canny-dev-lora`

Canny edge control LoRA from Black Forest Labs themselves. Provides structural guidance by conditioning on edge-detected input images.

`alvarobartt/ghibli-characters-flux-lora`

Studio Ghibli animation style for Flux. Captures the soft watercolor textures, rounded character designs, and warm palettes of Ghibli films.

Tips

Always read the model card before using a LoRA — it contains the trigger words and recommended settings.
Check the "Files and versions" tab to verify the repo contains .safetensors files.
If a LoRA has no visible effect, you likely forgot the trigger word or the base model doesn't match.
Trending LoRAs on HF are usually well-documented and tested by the community.