Skip to content

Make video

ziv-video makes clips with sound from a prompt or a starting image, and continues clips you already have. It needs ffmpeg. The guide Make video says what each setting does and why.

Make a clip

A video prompt has two parts: --scene says who and where, the look and the camera, and --prompt says what happens.

ziv-video -m ltx-4 --scene "A knight in a grey castle courtyard, overcast, handheld camera." --prompt "She draws her sword and turns to the gate."

You get a 5 second clip with sound, 768×448, saved as prompt_<date>_<time>.mp4 in the current folder. --prompt alone works too. The model sees the scene first, then the action; see Describe the scene and the action.

A vertical clip of about three seconds:

ziv-video -m ltx-4 --ratio 9:16 --size m --frames 73 --prompt "A cat walking through a garden, birds singing"

--frames is a frame count, not seconds; see How seconds map to frames. -m is ltx-4, ltx-8, ltx25-4 or h3; the models are compared in Models and licences.

Animate a still

ziv-video -m ltx-4 --image harbour.png --scene "A fishing harbour at dawn, calm water." --prompt "Gulls take off from the pier as a boat slides past."

The clip starts from harbour.png. The image is scaled to cover the video size and the overflow is cut from the centre, so pick a --ratio close to the image's shape. Say in the scene what the image shows and in the prompt what moves.

First and last frame

ziv-video -m ltx-4 --image start.png --end-image end.png --prompt "The lamp turns on"

The clip runs from one image to the other. A last frame needs a first frame, and every video model takes one. The clip ends close to the last image, not exactly on it, and two images of the same scene from about the same place work best. See First and last frame.

Follow an audio track

ziv-video -m ltx25-4 --image face.png --audio-file voice.wav --prompt "She talks to the camera"
ziv-video -m ltx-4 --audio-file song.mp3 --audio-start 90 --prompt "A band plays on a dark stage"

The picture is made to match the sound, and the clip's sound is exactly that track, so --no-audio has no effect. --audio-start is the second of the track where the clip starts. The clip keeps its length, and the sound is silence after the track ends. After each clip, the part of the track it used is printed, for example Audio track: song.mp3 1:30.0–1:34.0. A clear voice syncs best. See Follow an audio track.

Continue a clip (--extend)

ziv-video --extend clip.mp4 --prompt "She turns back to the sea."
ziv-video --extend clip.mp4 --extend-seconds 4 --prompt "She turns back to the sea."
ziv-video --extend clip.mp4 --extend-from 3.2 --prompt "She turns back to the sea."

You get a new file: the clip, followed by what happens next, with sound. The source is never changed. It uses the source's size, frame rate, scene, model and LoRAs unless you name others, so leave -m out. --prompt is what happens next and may be empty, which continues naturally.

  • --extend-seconds 4 sets the length to add (the default is 2), and --extend-frames 96 does the same in frames. The length is rounded to whole steps of 8 frames, or 17 with MiniMax H3.
  • --extend-from 3.2 continues from 3.2 seconds into the clip instead of its end: the frame there is the last one kept, and what follows it is dropped from the new file. The clip needs at least 25 frames before that point.
  • The size and length are the source's, so the options that choose them are refused; the list is in the command reference.
  • To carry a track on, extend with the same file: ziv-video --extend clip.mp4 --audio-file song.mp3. Without --audio-start, the extension starts where the source's sound ended.

See Extend a clip for what can be extended and how the scene and action carry over.

Use MiniMax H3

ziv-video -m h3 --prompt "A rider gallops across a salt flat at sunset, dust trailing behind" --frames 124 --steps 4

MiniMax H3 is stronger on dynamic scenes and much larger and slower than LTX. It has its own rules:

  • Frames are 17k+5: 22, 39, 56 … 124 (the default) … Another count moves up to the next one, with a warning.
  • Steps are 4 or 8; the default is 8. Another count moves to the nearest, a tie to 8.
  • The estimated time of the job is printed when it starts. It leaves out the one-time preparation of the first run.
  • Upscale (--upscale 2) works on Windows and Linux only, and not for a clip made from --audio-file.
  • Extend takes only 24 fps sources whose width and height are multiples of 32. Clips the app made pass.
  • --low-memory changes nothing: H3 always streams its blocks.
  • --lora takes MiniMax H3 LoRAs only.

Upscale while generating

ziv-video -m ltx-4 --upscale 2 --review --prompt "A sailing boat crossing a bay at golden hour"

The clip is rendered at half size, enlarged 2× and refined at full size. Width and height are rounded to 64 pixels so the half-size stage stays on the 32 pixel grid. --review pauses after each half-size render: press Y to upscale it, N to discard it or Q to quit. The first use downloads the upsampler on Windows and Linux. See Upscaling.

Use a LoRA or save memory

ziv-video -m ltx-8 --lora my-video-style:0.8 --prompt "A rainy street at night, neon signs"
ziv-video -m ltx-8 --no-low-memory --prompt "A rainy street at night, neon signs"

--lora works as for images: name:weight, comma-separated; see LoRAs. --low-memory is on for LTX and frees each part of the model once it has run; --no-low-memory keeps every part loaded, which needs more memory. ziv-video warns when a clip may not fit in memory and carries on. See Fit a clip in memory.

Run a prompt file

ziv-video -m ltx-4 -p clips.yaml -r 2

Every prompt of the file makes a clip, and scene: in an entry sets its scene. See Batches and prompt files.

How seconds map to frames

--frames counts frames, and both models run at 24 frames per second. Take seconds × 24 and the model moves it to a count it can make: LTX to the nearest 8k+1, MiniMax H3 up to the next 17k+5.

Length LTX MiniMax H3
2 s 49 56
3 s 73 73
5 s (the default) 121 124
8 s 193 192
10 s 241 243

A count that isn't valid is moved with a warning, so a close number works. There is no upper limit, but a longer clip needs more memory.

Flags

Every flag, with its default and limits, is in the command reference.

Flag Does In the Web UI
-m The model Model
--scene Who and where, the look, the camera Scene
--prompt What happens Action
-p A prompt file Prompt source → Prompt file
-r How many times to run each prompt Batch size
--ratio, --size Shape and resolution Ratio, Resolution
-W, -H Exact width and height W × H
--frames Length in frames Duration (s)
--steps Steps Steps
--seed The seed, used for every run Seed
--image, --end-image First and last frame First frame, Last frame
--audio, --no-audio Keep or remove the sound Audio
--audio-file, --audio-start An audio track to follow, and where in it to start Audio track, Start at
--extend Continue this video Extend, Extend from
--extend-from Continue from a point inside it Extend from
--extend-seconds, --extend-frames How much to add Extend by (s)
--lora LoRAs and their weights LoRAs
--upscale 2, --review Upscale while generating, and approve each small render Upscale, Review
--low-memory, --no-low-memory Free each part of the model once it has run Low memory
-o The output folder Output folder (Config page)
--format Output format; only mp4
--enhance, --enhance-note, --no-enhance Rewrite the prompt first ✨ Enhance, Direction
--enhance-model The enhancer model Prompt enhancer model (Config page)