Video generation parameters

A high-level overview of video generation parameters and when to use them. For parameters tied to reference images, keyframes, audio input, or video editing, see Capability-specific parameters at the bottom. For the complete schema, including every supported field along with its types and ranges, see the video generation API reference.

Available parameters vary by model: Wan 2.7 models use resolution and ratio instead of width and height. Kling requires keyframe images via media.frame_images instead of a prompt. See the supported models table for per-model coverage.

Quick reference

Match the problem you’re solving to the parameter most likely to help.

Video doesn’t match the prompt: Make the prompt more specific, add a negative_prompt for what to exclude, or raise guidance_scale toward 9-10.
Output looks oversaturated or has weird motion: Lower guidance_scale to 6-7. Avoid values above 12.
Poor visual quality: Raise steps to 30-40 for production runs. Diminishing returns past 50.
Generation is too slow or expensive while iterating: Lower steps to 10 for quick previews, and shorten seconds.
Need the same video every run (evals, regression tests): Set seed to a fixed integer.
Wrong dimensions or aspect ratio: Set width and height explicitly. On Wan 2.7, set resolution and ratio instead.
Output file is too large: Raise output_quality (higher number means more compression). Lower it for higher fidelity.
Need consistent characters or style across the video: Pass media.reference_images.
Need to pin starting or ending frames: Pass media.frame_images.
Need lip sync or beat-matched motion: Pass media.audio_inputs.

Prompting

`prompt`

A description of the video to generate. Required for every model except Kling. Maximum length is 32,000 characters. Be specific about subject, action, setting, camera movement, and pacing. Vague prompts produce generic motion. Include verbs and temporal cues (“slowly pans”, “the camera tracks left”) since video models are sensitive to motion language. Typical default: required.

`negative_prompt`

A description of what to avoid in the generated video. Useful for excluding common artifacts. Set it when the model produces unwanted elements (extra limbs, flickering, watermarks). A reasonable starting point: "blurry, low quality, distorted, flickering". Typical default: unset.

Output dimensions

`width` and `height`

The size of the generated video in pixels. Available combinations differ by model. Typical default: 1366 x 768.

`resolution`

A resolution tier used by Wan 2.7 models in place of width and height. Accepts "720P" or "1080P". Typical default: "1080P".

`ratio`

The aspect ratio used by Wan 2.7 models. Accepts "16:9", "9:16", "1:1", "4:3", or "3:4". Typical default: "16:9".

Length and frame rate

`seconds`

Clip duration in seconds. Accepted range is "1" through "10". Passed as a string. Longer clips cost more and take longer to generate. Use shorter clips while iterating on prompts and parameters. Typical default: "6".

`fps`

Frames per second. Higher values produce smoother motion at the cost of generation time and file size. Typical default: 24 (some models accept up to 60).

Quality and speed

`steps`

The number of denoising steps. More steps generally improve visual quality and temporal consistency at a near-linear cost in latency. Past a model-specific point, additional steps stop helping. Lower it (10) for quick previews. Use 20 for a balanced default. Raise it (30-40) for production runs. Avoid values above 50. Range: 10-50. Typical default: model-specific.

# Quick preview
job_quick = client.videos.create(
    prompt="A person walking through a forest",
    model="minimax/hailuo-02",
    steps=10,
)

# Production quality
job_production = client.videos.create(
    prompt="A person walking through a forest",
    model="minimax/hailuo-02",
    steps=40,
)

import Together from "together-ai";

const together = new Together();

// Quick preview
const jobQuick = await together.videos.create({
  prompt: "A person walking through a forest",
  model: "minimax/hailuo-02",
  steps: 10
});

// Production quality
const jobProduction = await together.videos.create({
  prompt: "A person walking through a forest",
  model: "minimax/hailuo-02",
  steps: 40
});

`guidance_scale`

Controls how closely the video follows the prompt. Higher values make the model adhere more strictly to the text description. Lower values give the model more creative freedom. Affects both visual content and temporal consistency. Recommended range is 6.0-10.0. Values above 12 may cause over-guidance artifacts or unnatural motion.

6.0-7.0: More creative, less literal.
7.0-9.0: Sweet spot for most use cases.
9.0-10.0: Strict adherence to the prompt.

Typical default: model-specific.

from together import Together

client = Together()

# Low guidance: more creative interpretation
job_creative = client.videos.create(
    prompt="an astronaut riding a horse on the moon",
    model="minimax/hailuo-02",
    guidance_scale=6.0,
    seed=100,
)

# High guidance: closer to literal prompt
job_literal = client.videos.create(
    prompt="an astronaut riding a horse on the moon",
    model="minimax/hailuo-02",
    guidance_scale=10.0,
    seed=100,
)

import Together from "together-ai";

const together = new Together();

// Low guidance: more creative interpretation
const jobCreative = await together.videos.create({
  prompt: "an astronaut riding a horse on the moon",
  model: "minimax/hailuo-02",
  guidance_scale: 6.0,
  seed: 100
});

// High guidance: closer to literal prompt
const jobLiteral = await together.videos.create({
  prompt: "an astronaut riding a horse on the moon",
  model: "minimax/hailuo-02",
  guidance_scale: 10.0,
  seed: 100
});

Reproducibility

`seed`

An integer that fixes the random initialization. With the same seed, prompt, model, and parameters, the model returns the same video. Useful for reproducibility and for fair comparisons when tuning other parameters. Typical default: unset (each call returns a new video).

Output format

`output_format`

The encoded video format. Accepts "MP4" or "WEBM". MP4 is the broadest-compatible default. WEBM produces smaller files but isn’t supported by every player. Typical default: "MP4".

`output_quality`

Compression quality. Lower values produce higher fidelity and larger files. Higher values produce smaller files with more compression artifacts. Typical default: 20.

Audio

`generate_audio`

Whether the model should generate audio for the video. Only applies to models that support audio generation. Typical default: false.

Capability-specific parameters

Parameter	Type	Description	Default
`prompt`	string	Text description of the video to generate.	Required
`model`	string	Model identifier.	Required
`width`	integer	Video width in pixels.	1366
`height`	integer	Video height in pixels.	768
`seconds`	string	Length of video (1-10).	`"6"`
`fps`	integer	Frames per second.	15-60
`steps`	integer	Diffusion steps (higher = better quality, slower).	10-50
`guidance_scale`	float	How closely to follow prompt.	6.0-10.0
`seed`	integer	Random seed for reproducibility.	any
`output_format`	string	Video format (MP4, WEBM).	MP4
`output_quality`	integer	Bitrate/quality (lower = higher quality).	20
`negative_prompt`	string	What to avoid in generation.	-
`resolution`	string	Video resolution tier (`720P`, `1080P`). Used by Wan 2.7 models instead of `width`/`height`.	`"1080P"`
`ratio`	string	Aspect ratio (`16:9`, `9:16`, `1:1`, `4:3`, `3:4`). Used by Wan 2.7 models.	`"16:9"`
`media`	object	Media inputs for the request (see schema and compatibility below).	-

These parameters belong to features with their own dedicated pages. Each link below covers supported models and end-to-end examples. The media object schema lists every field.

media.frame_images: Pin specific frames to known images (keyframes).
media.reference_images and media.reference_videos: Keep a character, object, or visual style consistent across the video.
media.audio_inputs: Drive generation with an audio file for lip sync, beat-matched motion, or narration.
media.source_video and media.frame_videos: Edit or extend an existing clip. Only Wan 2.7 models support these fields.

Pass frame_images and reference_images inside the media object. The SDKs don’t accept them as top-level parameters.

`media` object schema

The media object holds every image, video, and audio input for a video generation request. Not every model supports every field. For field compatibility across Wan 2.7 models, see the Wan 2.7 quickstart.

{
  "prompt": "...",
  "model": "...",
  "media": {
    "frame_images": [],
    "frame_videos": [],
    "reference_images": [],
    "reference_videos": [],
    "source_video": "",
    "audio_inputs": []
  }
}

Field	Type	Description
`frame_images`	array	Keyframe images for image-to-video models. Each item has the form `{input_image, frame}`, where `frame` is `"first"`, `"last"`, or a frame number. Without `frame`, one image becomes the first frame, two become the first and last frames, and three or more are spaced evenly from first to last.
`frame_videos`	array	Clips for video continuation on image-to-video models. Each item has the form `{video: "url"}`.
`reference_images`	array	Images for character or object consistency on reference-to-video models, or for visual guidance on video editing models.
`reference_videos`	array	Videos for character or object consistency on reference-to-video models. Each item has the form `{video: "url"}`.
`source_video`	string	URL of the video to edit, for video editing models.
`audio_inputs`	array	Audio file URLs that drive generation, such as lip sync or beat-matched motion, on text-to-video and image-to-video models. Each item is a URL string. Files must be WAV or MP3, 3 to 30 seconds long, and up to 15 MB.

GET STARTED

SERVERLESS

INFERENCE APIS

DEDICATED MODEL INFERENCE

DEDICATED CONTAINER INFERENCE

POST-TRAINING

FINE-TUNING

GPU CLUSTERS

CODE EXECUTION

ADMINISTRATION

Quick reference

Prompting

`prompt`

`negative_prompt`

Output dimensions

`width` and `height`

`resolution`

`ratio`

Length and frame rate

`seconds`

`fps`

Quality and speed

`steps`

`guidance_scale`

Reproducibility

`seed`

Output format

`output_format`

`output_quality`

Audio

`generate_audio`

Capability-specific parameters

`media` object schema

See also

​Quick reference

​Prompting

​prompt

​negative_prompt

​Output dimensions

​width and height

​resolution

​ratio

​Length and frame rate

​seconds

​fps

​Quality and speed

​steps

​guidance_scale

​Reproducibility

​seed

​Output format

​output_format

​output_quality

​Audio

​generate_audio

​Capability-specific parameters

​media object schema

​See also

Quick reference

Prompting

`prompt`

`negative_prompt`

Output dimensions

`width` and `height`

`resolution`

`ratio`

Length and frame rate

`seconds`

`fps`

Quality and speed

`steps`

`guidance_scale`

Reproducibility

`seed`

Output format

`output_format`

`output_quality`

Audio

`generate_audio`

Capability-specific parameters

`media` object schema

See also