LTX-2 license: models trained on its Outputs are derivatives
The LTX-2 Community License lists models trained on Outputs as Derivatives that must ship under the same license. What the text says and what it leaves open.
Yes, for LTX-2 itself: the license text defines "Derivatives of LTX-2" to include models created by generating synthetic data with LTX-2 to train another model, and clause 3(b) names "models trained on Outputs" as derivatives that must be distributed under the same agreement. If you plan to train on LTX-2 clips, read it before you build the dataset. This is not legal advice.
Quotes are from the LTX-2 Community License Agreement raw file on Hugging Face, read 2026-10-01. The copy I read ends at clause 7 and does not include Attachment A, so anything in that attachment is not covered here.
What does the license say about training on Outputs?
Three places matter. The definition of "Derivatives of LTX-2" covers any model "created or initialized by transfer of patterns of the weights, parameters, activations or output of LTX-2", including "methods based on the generation of synthetic data by LTX-2 for training the other model". Clause 3(b) says any Derivative, "including but not limited to fine-tuned weights, modified training code, models trained on Outputs", must be distributed exclusively under this agreement with a full copy of the license. Clause 5 says the licensor claims no rights in the Output you generate, "except as set forth herein".
Which parts apply to a dataset builder?
| Clause | What it says | What to check |
|---|---|---|
| Definition of Derivatives | Includes models trained on synthetic data generated by LTX-2 | Whether your pipeline fits that wording |
| 3(b) | Derivatives, including models trained on Outputs, ship only under this agreement | Your own model's license terms |
| 5 | Licensor claims no rights in Output, except as set forth in the agreement | The exceptions, including Attachment A |
Does the license apply to clips made elsewhere?
The license covers LTX-2 and Derivatives of LTX-2, and its "Output" means results of operating LTX-2. It does not describe other models. For a clip from any other model, the relevant terms are that model's own. Sume's docs do not state a training-on-outputs policy for any catalog model, so check the provider's terms for the model id you use. The ids, durations and resolutions Sume lists come from GET /v1/videos/models; see Video generation.
For the model itself, see LTX-2 open-source video model.
What should I keep if I build a dataset?
Record which model id produced each clip, the date, and the license text version you read. On Sume, video jobs accept an Idempotency-Key so a retry returns the original job instead of a second render, which keeps the record one-to-one with what you paid for. Billing is reserved on submit at provider list x 1.25.
Sources
Related posts
More in Developers
- LTX /v1/extend limits, and chaining clips with Sume frame_images
LTX's /v1/extend is 1080p only, adds 2-20 seconds, needs 73+ input frames. Sume has no extend call here: chain clips with frame_images, then join in Timeline.
- LTX retake vs Sume: fix one section of a video
LTX's POST /v1/retake regenerates one time region of a video. Sume's video edit takes a whole clip, so trim the bad section, edit it, and rejoin it.
- LTX frame counts must be 8k+1; Sume durations are whole seconds
LTX frame counts must satisfy (F-1) % 8 == 0, so 30 frames is invalid. On Sume you send whole seconds and read the allowed set from the catalog, not frames.
- Luma API 429 requests per minute: sliding window vs Sume
Luma counts requests in a sliding 60-second window and returns 429 if RPM or concurrent jobs fails. Sume returns 429 rate_limited: back off, reuse the key.
Written by Sume