LTX-2.5 model and workflow reference

LTX-2.5 is Sogni's default native video family: a 22B audio-video transformer that generates picture and synchronized 48 kHz stereo audio in a single pass. Fast, HQ, and Pro currently route through the release-validated official Distilled/Turbo workflows. Dev checkpoints are not publicly routed until upstream publishes an official ComfyUI Dev recipe and Sogni validates it.

Every public mode runs a fixed 8-step distilled path at guidance 1, so an LTX-2.5 render of a given shape costs and takes roughly what any other LTX-2.5 render of that shape does.

For a mode-by-mode walkthrough with side-by-side comparisons against MiniMax H3 and LTX-2.3, see H3 Directs. LTX-2.5 Brings the Camera Rig. on the Sogni blog.

#Workflow IDs

Workflow	Creative Agent selector	Public model ID	Inputs
Text to video	`ltx25`	`ltx25-22b-int8_t2v_distilled`	Prompt
Image to video	`ltx25`	`ltx25-22b-int8_i2v_distilled`	Prompt + first frame
First/last-frame video	`ltx25` with both frame roles	`ltx25-22b-int8_i2v_distilled`	Prompt + first and last frames
Audio to video	`ltx25-a2v`	`ltx25-22b-int8_a2v_distilled`	Prompt + audio
Image + audio to video	`ltx25-ia2v`	`ltx25-22b-int8_ia2v_distilled`	Prompt + first frame + audio
Video to video	`ltx25-v2v`	`ltx25-22b-int8_v2v_distilled`	Prompt + source video (+ mask for inpaint)

First/last-frame generation uses a dedicated FLF workflow template while sharing the public I2V model ID. There is no public Dev selector.

#Generation limits

Resolution 640–3840 px per side on a 64 px grid, default 1920×1088. Video-to-video uses a 128 px grid, default 1920×1024.
Frames 25–505 on a 1 + 8n grid, default 97 — roughly 1–21 seconds at 24 fps. Sogni's current public workflows are exercised up to 20 seconds.
Frame rate is yours to set; 24 fps is the cinematic default.
Audio 48 kHz stereo, generated jointly with the picture rather than dubbed on afterwards.
Steps fixed at 8 for every public distilled mode; guidance fixed at 1.

Resolution and frame rate trade against each other at the top of the range: 4K is available, and 60 fps is available at 1080p, but not both at once.

#Pricing

LTX-2.5 is priced from the shape of the job — pixels × frames — rather than a flat per-second rate, so a short 4K clip and a longer 1080p clip can land at similar cost. Measured examples from Sogni reference runs on fast-network workers:

Mode	Resolution	Length	Spark	USD
Multishot text to video	1920×1088	10 s	91.83	$0.46
Image to video	1088×1536	14 s	102.72	$0.51
First/last-frame video	1920×1088	8 s	73.54	$0.37
Video to video (canny)	1920×1024	8 s	69.14	$0.35

Prices are live network estimates at the standard 1 Spark = $0.005 peg. For a quote at any shape, use the job estimate calculator. Every LTX-2.5 render is also covered credit-free under fair use on Sogni Unlimited plans.

#Render time

LTX-2.5 runs on Sogni Supernet fast-network workers, whose GPUs differ, so wall-clock time varies by job. In Sogni reference runs an 8-second 1080p render ranged from about 1 m 20 s to over 3 m depending on the worker. All public modes run 8 steps, against 30 for the Dev tier.

#Video-to-video controls

The LTX-2.5 V2V family supports canny, depth, pose, detailer, inpaint, and outpaint templates. Pose requires a source video for the motion sequence and a still reference image for the subject's appearance. Inpaint also requires a mask.

#Prompt shape

Use one continuous natural-language scene paragraph with concrete subject, action, setting, lighting, camera motion, and synchronized audio cues.

Put spoken dialogue in quotation marks, and name the language and accent when it matters.
Break long speeches into short fragments with acting directions between them rather than one unbroken block — this is upstream LTX guidance and it materially changes delivery.
State the camera on every shot, including when it should not move. Leaving it unstated invites drift in the final seconds.
Keep the amount of speech realistic for the requested duration.
For image to video, treat the still as visual truth: describe motion, performance, camera and sound, not the static contents the model can already see.
For first/last-frame video, describe the motion and change between the supplied anchors instead of re-inventing their static contents.

#Multi-shot prompting

LTX-2.5 will cut between shots inside a single generation. To get a real hard cut rather than a dissolve, give each shot its own sentence, state a numeric timestamp for the cut, and make the shots genuinely different places. A change of room with a change of light reads as an edit; a tighter framing of the same subject usually resolves as a push-in instead. Keep multi-shot captions to roughly one sentence per shot.

#Sample gallery

Text to video, 1920×1088, 10 s. Three shots and two hard cuts from a single prompt, with synchronized audio — no editing.

Image to video, 1088×1536, 14 s. A macro portrait animated from a single still, holding skin texture and identity through a spoken line.

First/last-frame video, 1920×1088, 8 s. Two anchored stills of the same summit, sunrise to deep night, with nothing moving but the light.

More comparisons, including matched A/B pairs against LTX-2.3 and MiniMax H3, are in the LTX-2.5 launch post.

#LTX-2.3 rollback boundary

LTX-2.3 remains available through explicit rollback selectors. Voice ID-LoRA, the community transition LoRA, and 10Eros remain LTX-2.3-only; do not send those inputs or triggers to an LTX-2.5 model ID.