LTX-2.5 Cinemagraph LoRA vs image-to-video on Sume
The LTX-2.5 Cinemagraph LoRA is an image-to-video adapter for selective motion. On Sume, start from first_frame and describe the motion in the prompt.
The LTX-2.5 Cinemagraph LoRA is an adapter with an install snippet for local LTX-2 pipelines, tagged image-to-video and selective-motion; the page header shows no hosted endpoint. To get a similar loop on Sume, submit a still as first_frame on a video model and write the motion you want, and what you want to stay still, in the prompt.
Model facts are from the Hugging Face page, read 2026-10-01. The snapshot I could read is the page header only: the model card text and files were not available, so I make no claims about how the LoRA behaves or how to prompt it. Sume facts are from Video generation.
What does the model page show?
The page lists the model as Lightricks/LTX-2.5-22b-LoRA-Cinemagraph, pipeline tag Image-to-Video, tags cinemagraph, i2v, lora, ltx-2.5, motion-design and selective-motion, and the license ltx-2.x-community-license. Its install snippet clones the LTX-2 repository and downloads the adapter weights; base components come from Lightricks/LTX-2.5.
What does Sume offer for selective motion?
There is no LoRA input. The video catalog advertises frame image types and reference types per model, and the docs advise including "details about motion, camera angles, lighting, and scene composition" in the prompt. Selective motion is therefore a prompting job: name the element that moves and say the rest of the frame holds still.
| Step | LTX-2.5 Cinemagraph LoRA | Sume video request |
|---|---|---|
| Start image | Image-to-video pipeline | first_frame where the model lists it in supported_frame_images |
| Motion control | Adapter weights | Motion described in the prompt |
| Length | Not stated in the header snapshot | Per model; minimax-h3 lists 5-15 s, most others stop at 15 s |
| Where it runs | Your machine | Hosted job; retries safe with Idempotency-Key |
How do I check a model supports first_frame?
Read GET /v1/videos/models and look at supported_frame_images for the model id; the docs example shows ["first_frame", "last_frame"]. A loop that ends where it began can use the same still as both frames on a model that lists both.
For a no-code walkthrough of the same idea, see cinemagraph AI from a photo.
Will the result be a true cinemagraph?
Not guaranteed. A prompt can ask for a still background, but the model decides. Run a short clip first, compare the frozen regions, and adjust the wording before you spend on a longer render. Billing is reserved on submit at provider list x 1.25.
Sources
Related posts
More in Models
- Luma layers API: 10 RGBA layers vs Sume's flat image result
Luma type layering splits one image into up to 10 ordered RGBA PNG layers. Sume returns flat images at data[].url; n counts images, not layers.
- Ray 3.2 API aspect ratios and resolutions, and Sume's lists
Luma lists six Ray 3.2 aspect ratios (9:16 to 21:9) at 540p, 720p and 1080p. Sume reports ratios per model in supported_aspect_ratios and rejects size.
- Luma Scenes Ray 3.2 or Seedance 2: pin a model or use auto
Luma says Ray 3.2 and Seedance 2 are not interchangeable. On Sume, pin a catalog model for a known clip length, or send sume/auto and let Sume pick.
- Seedance 2.0 4K price: Luma 261 credits/sec, Runway 150, Sume 1080p
Luma lists Seedance 2.0 at 4K for 261 credits per second; Runway's docs list 150. Sume's docs list seedance-2 at 480p, 720p and 1080p, with no 4K entry.
Written by Sume