Cinema Set Freecinemasetfree.com
All LoRA Guides

Using Replicate LoRAs

Replicate models accept LoRAs through the hf_lora parameter (for loading external LoRAs at runtime) or as trained fine-tunes you run directly. The lora_scale parameter controls strength (0.0–1.0). You can combine two LoRAs using extra_lora and extra_lora_scale, or use multi-LoRA models for up to 20 simultaneous LoRAs.

Two Ways to Use LoRAs

Approach A: Trained Fine-Tunes (Replicate-Hosted)

When you train a LoRA on Replicate (using replicate/fast-flux-trainer or ostris/flux-dev-lora-trainer), the result is a runnable model at yourusername/your-model-name. You run it like any other Replicate model — the LoRA is already baked into the pipeline.

Include the trigger word in your prompt to activate the trained concept. The lora_scale parameter (default 1.0, range −1.0 to 2.0) controls the primary LoRA strength.

Training typically costs under $2 and takes 2–20 minutes depending on the trainer and dataset size. Fast-flux-trainer is the quickest option, while ostris/flux-dev-lora-trainer offers more configuration options for advanced users.

Approach B: Runtime LoRA Loading (hf_lora)

Models like black-forest-labs/flux-dev-lora accept an hf_lora parameter that loads any compatible LoRA at inference time. Pass a Hugging Face repo ID, CivitAI download URL, or a direct link to a .safetensors file.

The model downloads and caches the LoRA weights, then applies them during generation. The lora_scale parameter (default 0.8, range 0.0–1.0) controls how strongly the LoRA affects the output. This approach is ideal for experimenting with community LoRAs without training anything yourself.

Supported URL Formats

Hugging Face

Format: owner/model-name

Example: alvdansen/frosting_lane_flux

Replicate

Format: owner/model-name

Example: jakedahn/flux-midsummer-blues

CivitAI

Format: Full download URL

Example: https://civitai.com/api/download/models/735262?type=Model&format=SafeTensor

Direct File

Format: URL to .safetensors file

Example: https://example.com/lora.safetensors

Key Parameters

hf_lora (string)

The LoRA to load at runtime. Accepts Hugging Face repo IDs, Replicate model IDs, CivitAI download URLs, or direct links to .safetensors files. The model will download and cache the LoRA weights on first use.

lora_scale (float, 0.0–1.0)

How strongly the primary LoRA affects the output. Default is 0.8 for runtime-loaded LoRAs and 1.0 for trained fine-tunes. Start at 0.8 and adjust based on results.

extra_lora (string)

A second LoRA to combine with a trained fine-tune. Accepts the same URL formats as hf_lora. Useful for pairing a character LoRA (primary) with a style LoRA (extra).

extra_lora_scale (float, −1.0 to 2.0)

Strength for the second LoRA. Wider range than the primary lora_scale, allowing you to amplify or even negate a LoRA's effect. Negative values invert the learned features.

Models with LoRA Support

black-forest-labs/flux-dev-lora

Official BFL Flux Dev with LoRA support. The recommended model for most use cases — high quality output with reliable LoRA loading. Supports hf_lora, extra_lora, and lora_scale parameters.

black-forest-labs/flux-schnell-lora

Schnell variant optimized for speed. Generates images in 1–4 steps instead of 20–50. Trades some quality for significantly faster generation, making it ideal for rapid iteration and testing LoRAs.

lucataco/flux-dev-multi-lora

Supports up to 20 simultaneous LoRAs via the hf_loras array parameter (note the plural). Each LoRA can have its own weight via the lora_scales array. Best for complex multi-concept compositions.

Trained Fine-Tunes

Any model trained via Replicate's fine-tuning trainers (fast-flux-trainer, ostris/flux-dev-lora-trainer) automatically gets its own runnable endpoint with LoRA support, including extra_lora for combining with a second adapter.

Tips

  • Always include the LoRA's trigger word in your prompt — without it, the LoRA may have no visible effect on the generated image.
  • Start lora_scale at 0.8 and adjust from there. Going above 1.0 can oversaturate the LoRA's effect, causing artifacts and reduced prompt adherence.
  • CivitAI URLs require an API token appended as ?token=YOUR_TOKEN for gated or early-access models.
  • For multi-LoRA setups, keep total combined weights under 2.0 to avoid quality degradation. A character LoRA at 0.9 + a style LoRA at 0.5 is a good starting point.
  • Flux-based LoRAs only work with Flux models. SDXL LoRAs only work with SDXL. They are not cross-compatible — using the wrong base model will produce garbage output or errors.
  • The first run with a new hf_lora value will be slower as Replicate downloads and caches the LoRA weights. Subsequent runs with the same LoRA are faster.