LoRAs | LTX Documentation

A LoRA (Low-Rank Adaptation) is an adapter that changes how an LTX model generates without replacing the base model. Use a standard LoRA when you want a repeatable visual style, appearance, effect, or learned motion behavior.

For reference-driven control, restoration, VFX, or creative transforms, use an IC-LoRA instead.

Choose the right LoRA

If you need to…	Use	Why
Apply a visual style, appearance, effect, or prompt-driven behavior	Standard LoRA	The LoRA modifies the model’s generation behavior.
Follow a reference video, depth map, pose, edges, or motion path	IC-LoRA	The LoRA uses a reference input to guide generation.
Use both	Both, only when the LoRA cards and workflow establish compatibility	Each LoRA has its own model and workflow requirements.

Before you begin

You need:

A current ComfyUI installation with ComfyUI-LTXVideo. See Using ComfyUI with LTX for setup and model files.
A standard LoRA that is compatible with your LTX model and workflow.
The LoRA’s model card or README. It is the source of truth for its compatible base model, required prompt wording, loader, and recommended strength.

The LTX-2.5 example-workflow directory includes task-specific IC-LoRA workflows, but it does not currently include a general-purpose standard-LoRA workflow. Using a standard LoRA requires a manual adaptation to the baseline template.

Use a standard LoRA in ComfyUI

1. Check the LoRA card

Before downloading a LoRA, confirm all of the following:

It is a standard LoRA, not an IC-LoRA or another specialized LoRA type.
Its supported LTX version and base transformer match the model you plan to use.
It has any required trigger words or prompt guidance.
Its card specifies a loader or strength recommendation.

Do not assume that a LoRA trained for one checkpoint, quantization, or workflow will behave the same way in another.

2. Choose a workflow

The ComfyUI guide lists available templates from LTX and their model files. Load the template that matches your task, then run it once unchanged. This confirms that the base workflow and model stack work on your system.

Base workflow	Standard-LoRA status	Required adaptation
First-Frame / Last-Frame (single-stage)	Manual adaptation required	Insert the model-only loader between the template’s shared Model output and its sampler.
Text-to-Video or Image-to-Video (two-stage)	Manual adaptation required	Insert one model-only loader after the shared Model output, then connect its output to both the Stage 1 and Stage 2 sampler inputs. Do not apply it to only one stage.
Audio-to-Video two-stage workflow	Manual adaptation required	Use the same shared-model wiring for both stages. Keep the workflow’s audio-conditioning path intact, and make sure the LoRA is compatible with the A2V workflow.
IC-LoRA control or editing workflow	No default standard-LoRA adaptation	Combine only when the relevant LoRA cards and workflow explicitly establish compatibility.

The workflow and asset reference identifies which official workflows are single-stage or two-stage.

3. Install the LoRA

Copy the LoRA’s .safetensors file to:

ComfyUI/models/loras/

Restart ComfyUI, or refresh its model list, so the file appears in the loader.

4. Apply the LoRA to the model

For a standard LTX LoRA, add ComfyUI’s core LoraLoaderModelOnly node. It accepts a MODEL, applies the selected LoRA at strength_model, and returns the modified MODEL. Select your LoRA file and use the wiring in the table above.

Do not use the general LoraLoader node as the default replacement: it requires both a MODEL and a CLIP input and attempts to modify both. Use it only when the LoRA card specifically calls for it. The model-only loader is the appropriate starting point for a LoRA that modifies the LTX diffusion model.

5. Test a controlled first run

Use the prompt wording and strength from the model card. Keep the seed, prompt, resolution, frame count, and other generation settings fixed while you test the LoRA. This makes it possible to see what the LoRA changed.

Set LoRA strength

The LoRA’s model card takes precedence over any generic value. Different LoRAs are trained and scaled differently.

If a card does not give a starting value, test at 1.0 first, then change only the strength:

Lower it when the LoRA overwhelms the prompt, removes too much of the base model’s behavior, or introduces artifacts.
Raise it only when the intended effect is too weak and the LoRA card does not advise a different approach.
Re-run with the same seed and prompt after each change.

Do not treat a strength that works for one LoRA as a recommendation for another. In particular, IC-LoRA strength guidance applies only to the specific IC-LoRA workflow and adapter.

Combine LoRAs carefully

First verify each LoRA on its own. To combine compatible standard LoRAs, chain LoraLoaderModelOnly nodes, then connect the final modified MODEL output to every sampler stage in the workflow. Check every LoRA card for compatibility requirements before combining them.

Avoid universal strength caps. The useful range depends on the LoRAs, their training, the base model, and the workflow.

Compatibility and performance

The LTX-2.5 model card states that a large majority of LTX-2.3 LoRAs and IC-LoRAs run unchanged on LTX-2.5, with some exceptions. Validate the exact LoRA and workflow before relying on it in production.

Quantized model stacks and LoRAs must also be tested together. Do not assume a LoRA works with every quantization or loader just because it works with the BF16 model stack.

A LoRA adds model weights and can affect model-loading time and memory use. The exact cost depends on the LoRA, your chosen model stack, and how the workflow manages the weights. Test your target resolution and duration on the hardware you plan to use.

Troubleshooting

The LoRA has no visible effect

Confirm that the workflow is using the modified MODEL output from the LoRA loader — in a two-stage workflow, connect it to both the Stage 1 and Stage 2 samplers, not just one.
Check that the LoRA matches the selected LTX model and transformer.
Use the LoRA’s required trigger words or prompt structure.
Test the card’s recommended strength with a fixed seed.

The LoRA dominates the output or creates artifacts

Lower the LoRA strength and retest with the same seed.
Simplify the prompt so it does not conflict with the LoRA’s learned behavior.
Check the model card for its intended resolution, prompt format, and model pairing.

The LoRA does not appear in ComfyUI

Confirm that the .safetensors file is in ComfyUI/models/loras/.
Refresh the model list or restart ComfyUI.
Check the ComfyUI console for a loader error or a missing custom-node dependency.

Train a custom LoRA

Use the LTX Trainer to train a custom LoRA. The trainer documentation owns the supported training configurations; see LTX training for the entry point.