LTX-2.3 LoRAs on LTX-2.5: most run unchanged; Sume has no LoRA field

Most of them, according to Lightricks. The LTX-2.5 model card says that, based on its testing, the large majority of LoRAs and IC-LoRAs trained on LTX-2.3 run on LTX-2.5 without changes, that a small number of exceptions exist, and that you should validate your adapters before production use. Sume cannot load a LoRA at all: its video request schema has no such field, and model_params has an empty allowlist.

The LTX statements are from the LTX-2.5 model card, read 2026-10-02. The Sume statements are from the OpenAPI schema and Video Router.

What does Lightricks say about LoRA compatibility?

The card's training section gives two statements. The dev transformer is fully trainable, and you can reproduce published LoRAs and IC-LoRAs with the LTX-2 Trainer. Separately, adapters trained on LTX-2.3 mostly carry over, with a small number of exceptions that the card does not name.

The card does not give a percentage, and it does not say which adapters break. The practical reading is a test, not a guarantee: run each adapter you depend on against a fixed prompt set before you swap models.

How do I validate an adapter after switching to LTX-2.5?

Keep five to ten prompts and first frames that you already have reference output for. Render each with the adapter on LTX-2.5 using the same seed and settings, and compare identity, motion and text legibility against the LTX-2.3 render. An adapter that changes character likeness or drops its trigger behavior is one of the exceptions the card warns about.

Do this on the distilled and the dev checkpoint if you use both, because the card lists them as different files.

Can I use a LoRA through the Sume API?

No. The Video Router request schema sets additionalProperties to false and lists fields such as model, prompt, image_url, reference_image_urls, resolution, duration and model_params. There is no LoRA field, and model_params is documented as reserved, empty allowlist in v1, so omit it or send {}.

If you need a style fixed across clips on Sume, the documented levers are reference images (reference_image_urls) on the models that accept them, and a first frame for image-to-video. Check each model's supported_input_references in the catalog.