Cinema Set Freecinemasetfree.com
All LoRA Guides

Using Replicate LoRAs

CSF's image models run on Replicate. To use a Replicate LoRA, open the LoRA / Integrations modal on a design and paste the model as owner/model (or owner/model/version to pin a version) into a LoRA Weights field, then set the LoRA Scale slider for strength. Some models add an Extra LoRA field so you can stack a second adapter.

Where LoRAs Live in CSF

On a project's Designs tab, expand a design and open its image settings, then click LoRA Integrations. This button only appears when the design's image model exposes LoRA parameters (a model with a lora_weights or lora_scale input), so pick a LoRA-capable Image Model first if you don't see it.

The button opens the LoRA / Integrations modal, whose model picker is pre-set to the design's image model. The fields below are that model's LoRA and integration parameters. Set your values and click Save — they're stored as a preset for the design and applied every time it generates images. Reset to Defaults restores the model's defaults; Cancel discards changes.

Two Ways to Use Replicate LoRAs

A: Reference a Trained Fine-Tune

When you train a LoRA on Replicate (using a trainer like replicate/fast-flux-trainer or ostris/flux-dev-lora-trainer), the result is a runnable model at yourusername/your-model-name. Training happens on Replicate, outside CSF, and typically costs a couple of dollars and takes a few minutes.

To use it, paste that model ID into a LoRA Weights field in the modal — add /version if you want to lock a specific version. Include the LoRA's trigger word in your design's prompt to activate the trained concept.

B: Load a Community LoRA

LoRA-capable Flux models let you load any compatible LoRA into a LoRA Weights field without training anything. The reference can be a Replicate model ID, a Hugging Face repo, a CivitAI model page, or a direct .safetensors URL — the modal lists these formats as hints under the input.

On its first use the LoRA is downloaded and cached, so that generation is a little slower than later ones. The LoRA Scale slider controls how strongly the LoRA affects the output; its range and default come from the model you picked.

Accepted Reference Formats

Every LoRA field in the modal accepts the same four formats, shown as hints beneath the input:

Replicate

Format: owner/model or owner/model/version

Example: jakedahn/flux-midsummer-blues

Hugging Face

Format: huggingface.co/owner/model

Example: huggingface.co/alvdansen/frosting_lane_flux

CivitAI

Format: model page URL

Example: civitai.com/models/12345

Direct File

Format: URL to a .safetensors file

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

Not sure where to look? The modal has Browse Replicate LoRAs, Browse CivitAI LoRAs, and Browse HuggingFace LoRAs links that open each source's LoRA listing in a new tab.

Fields You'll See

The exact controls depend on the model you pick. Common ones include:

LoRA Weights (text)

The primary LoRA to load, entered in one of the four formats above. The model downloads and caches the weights on first use.

Extra LoRA (text)

A second LoRA to stack on top of the primary one, using the same reference formats. Handy for pairing a character LoRA with a style LoRA. Only appears on models that support it.

LoRA Scale (slider)

How strongly the LoRA affects the output. The slider's minimum, maximum, and default are read from the model's schema, and the current value is shown next to the label. Start near the model's default and adjust from there.

Fine-tune ID (text)

Some hosted fine-tune models take an ID for a previously trained fine-tune instead of a weights URL. When present, paste the fine-tune's ID here.

Tips

  • Always include the LoRA's trigger word in your design's prompt — without it, the LoRA may have no visible effect on the generated image.
  • Start the LoRA Scale near the model's default and adjust from there. Pushing it toward the top of the slider's range can oversaturate the effect and reduce prompt adherence.
  • Flux-based LoRAs only work with Flux models; SDXL LoRAs only work with SDXL. They are not cross-compatible — using the wrong base model produces garbage output or errors.
  • The first generation with a new LoRA reference is slower while the weights download and cache; later runs with the same reference are faster.
  • CivitAI download links for gated or early-access models may require a token appended as ?token=YOUR_TOKEN.
  • Once you're happy, Save to store the configuration as the design's preset so you don't have to re-enter it.