Skip to main content

Image2Video

Image2Video

See example prompts in the gallery.

Pricing for video generation.

Overview​

Astria Image2Video supports a wide catalog of image-to-video, text-to-video, reference-to-video, and motion-control video models. Pick a video_model family, optionally a video_resolution, and pass a video_prompt describing the motion; Astria renders the first frame using the prompt's image stage (the chosen tune + text) and animates it.

Form fields​

Pass these alongside text on POST /tunes/:id/prompts:

video_model (required)​

The model family to animate with, for example seedance25 or veo31_fast. See Model families below. Legacy variant ids such as seedance25_720p or veo31_fast_audio_1080p are still accepted; see Variant ids.

video_resolution (optional)​

Output resolution within the family, for example 720p or 1080p. When omitted, the family's default resolution is used. Families without a resolution choice take no video_resolution; sending one returns a 422. With a legacy variant id, an explicit video_resolution replaces the id's own resolution within its family.

video_prompt (required)​

Describes camera movement, scene, and object interactions. Keep image-stage LoRA tokens in text. For Seedance 2 reference-to-video, put reference tune tokens directly in video_prompt using <faceid:TUNE_ID:1> TUNE_NAME (or the corresponding lora token) so Astria can supply those tunes' images to the video model.

Example: <lora:1533312:1.0> ohwx woman hiking in the alps (in text) + Woman looking at the camera, smiling, puts hands on her hips, confident (in video_prompt).

image_references (optional)​

Ordered array of multipart image uploads for reference-to-video models. Repeat prompt[image_references][] for each image.

image_reference_urls (optional)​

Ordered array of public image URLs, as an alternative to image_references. Repeat prompt[image_reference_urls][] for each URL. Order is preserved within each array.

video_duration (optional)​

Integer seconds. Allowed values depend on the chosen model — see the table.

video_first_frame (optional)​

Multipart image upload. When provided it overrides the image-stage render and text is no longer required.

video_first_frame_url (optional)​

URL alternative to video_first_frame.

video_last_frame (optional)​

Multipart image upload for first+last keyframe models.

video_last_frame_url (optional)​

URL alternative to video_last_frame.

source_image_url (optional)​

URL of an existing generated image to animate. This skips the image-generation stage. An explicit video_first_frame takes precedence.

video_references[] (optional)​

Ordered array of source or reference videos. Repeat prompt[video_references][] for each one; each item can be a multipart upload, a signed upload ID, or a public URL, in any mix. The prompt names them by position as @Video1, @Video2. Motion-control models require one as their driving video, and Edit models take one as their source.

audio_references[] (optional)​

Ordered array of reference audio tracks, supplied like video_references[] and named @Audio1, @Audio2. Attaching one sets video_audio=true. Astria adds the token on provider paths that require it.

Seedance 2.5 accepts up to 10 videos and 10 audio tracks in one request, each 2–30 seconds, with at most 30 seconds in total for videos and a separate 30 seconds for audio. Every other model accepts one of each. max_reference_videos and max_reference_audios in the live catalog give each model's limit.

input_video, input_video_url, audio_reference, audio_reference_url (legacy)​

Still accepted, as the first item of the matching array. Do not send a legacy field together with its array. Responses keep input_video and audio_reference as the first item of video_references and audio_references.

video_audio (optional)​

Boolean. Generate an audio track when supported by the model. Defaults to false. It is set automatically when audio_references[] is attached. For families that price audio separately (marked Toggle in Model families), video_audio=true selects the audio variant, for example veo31_fast + 1080p + audio renders veo31_fast_audio_1080p.

aspect_ratio (optional)​

Forwarded to the video model when it supports an aspect-ratio knob. Seedance 2 accepts 16:9, 9:16, 1:1, 4:3, 3:4, 21:9, or adaptive. For image-to-video the aspect ratio is derived from the input image.

Example​

curl -X POST -H "Authorization: Bearer $API_KEY" \
https://api.astria.ai/tunes/$TUNE_ID/prompts \
-F 'prompt[text]=<lora:1533312:1.0> A highly detailed image of thoughtful ohwx woman exploring a hidden urban garden' \
-F 'prompt[video_model]=seedance_v15_720p' \
-F 'prompt[video_prompt]=Woman looking at the camera, smiling, puts hands on her hips, confident' \
-F 'prompt[video_duration]=5' \
-F 'prompt[aspect_ratio]=16:9'

Family and resolution​

Address a model as a family plus a separate video_resolution. This prompt renders Seedance 2.5 at 1080p; omit video_resolution to use the family's default (720p).

curl -X POST -H "Authorization: Bearer $API_KEY" \
https://api.astria.ai/tunes/$TUNE_ID/prompts \
-F 'prompt[video_model]=seedance25' \
-F 'prompt[video_resolution]=1080p' \
-F 'prompt[video_prompt]=A lone explorer walks across endless dunes at sunrise, cinematic.' \
-F 'prompt[video_duration]=5' \
-F 'prompt[aspect_ratio]=16:9' \
-F 'prompt[video_audio]=true'

The prompt response keeps video_model as the resolved variant id (seedance25_1080p) and adds video_model_family (seedance25) and video_resolution (1080p).

Text-to-video​

Models in the text-to-video group accept a video_prompt with no first frame and no text — the model generates the video directly. Currently: seedance2_*, happyhorse_720p, happyhorse_1080p.

curl -X POST -H "Authorization: Bearer $API_KEY" \
https://api.astria.ai/tunes/$TUNE_ID/prompts \
-F 'prompt[video_model]=seedance2_fast_720p' \
-F 'prompt[video_prompt]=A lone explorer walks across endless dunes at sunrise, cinematic.' \
-F 'prompt[video_duration]=5' \
-F 'prompt[aspect_ratio]=16:9' \
-F 'prompt[video_audio]=true'

Seedance 2 reference media​

Seedance 2 can condition a video on multiple tune images, one reference video, and one audio reference; Seedance 2.5 takes up to 10 of each. Mention existing tunes in video_prompt using the same token and class-name syntax as image generation. Up to nine reference images are sent in total (up to three per tune).

curl -X POST -H "Authorization: Bearer $API_KEY" \
https://api.astria.ai/tunes/$TUNE_ID/prompts \
-F 'prompt[video_model]=seedance2_fast_720p' \
-F 'prompt[video_prompt]=<faceid:1234:1> woman walks onto the stage in time with @Audio1 while the camera follows' \
-F 'prompt[video_duration]=8' \
-F 'prompt[aspect_ratio]=16:9' \
-F 'prompt[video_references][]=@/path/to/movement-reference.mp4' \
-F 'prompt[video_references][]=@/path/to/camera-reference.mp4' \
-F 'prompt[audio_references][]=@/path/to/music-reference.mp3' \
-F 'prompt[video_audio]=true'

Use a public URL as the item instead when the file is already publicly accessible. Do not combine Seedance 2 or Seedance 2.5 first/last frames with image_references, image_reference_urls, video_references, or audio_references; these are separate media modes.

Ordered image references​

Use image_references when each frame should be conditioned directly by a sequence of images rather than tune training images. The same field works with any compatible reference-to-video model; model-specific image limits apply.

curl -X POST -H "Authorization: Bearer $API_KEY" \
https://api.astria.ai/tunes/$TUNE_ID/prompts \
-F 'prompt[video_model]=seedance2_fast_720p' \
-F 'prompt[video_prompt]=Begin on the back detail, orbit around the model, and end on the front view' \
-F 'prompt[video_duration]=8' \
-F 'prompt[image_references][]=@/path/to/back.jpg' \
-F 'prompt[image_references][]=@/path/to/front.jpg'

For hosted images, send repeated prompt[image_reference_urls][] fields instead. The prompt response exposes the stored references as an ordered image_references array of URLs. On models with exclusive media modes, image references cannot be combined with first/last frames.

Motion control​

Motion-control models replay the motion from video_references[] while restyling the subject. Models: kling30_motion_control, kling30_motion_control_pro, wan_animate_720p, dreamactor_m2, happyhorse_motion_control.

curl -X POST -H "Authorization: Bearer $API_KEY" \
https://api.astria.ai/tunes/$TUNE_ID/prompts \
-F 'prompt[text]=ohwx man <faceid:123>' \
-F 'prompt[video_model]=kling30_motion_control_pro' \
-F 'prompt[video_prompt]=match the dance moves' \
-F 'prompt[video_duration]=10' \
-F 'prompt[video_references][]=@/path/to/reference.mp4'

Model families​

The current families, generated from video_model_families in the live catalog at GET https://api.astria.ai/cli/models. Fetch the catalog for the up-to-date list and each family's capability flags (supports_last_frame, supports_image_references, supports_audio_reference, needs_input_video, generates_audio, native_audio, exclusive_media_modes, aspect_ratios).

  • video_resolution: the values the family accepts. — (none) means the family has no resolution choice and rejects video_resolution with a 422.
  • Default: the resolution used when video_resolution is omitted.
  • video_audio: Toggle families price audio separately, so video_audio=true picks the audio variant. Other families that generate audio pass video_audio to the model as generate_audio.
video_model familyTitlevideo_resolutionDefaultvideo_audio
seedance25Seedance 2.5480p, 720p, 1080p720pPassed to the model as generate_audio
seedance25_draftSeedance 2.5 Draft 480p— (none)—Passed to the model as generate_audio
wan30WAN 3.0480p, 720p, 1080p720pPassed to the model as generate_audio
wan30_editWAN 3.0 Edit480p, 720p, 1080p720pPassed to the model as generate_audio
wan30_prime_editWAN 3.0 Prime Edit480p, 720p, 1080p720pPassed to the model as generate_audio
flux3FLUX 3720p, 1080p720pPassed to the model as generate_audio
minimax_h3_maxMiniMax H3 Max480p, 768p480pNative audio
minimax_h3MiniMax H32k2kNative audio
seedance25_editSeedance 2.5 Edit480p, 720p, 1080p720pPassed to the model as generate_audio
seedance2Seedance 2720p720pPassed to the model as generate_audio
seedance2_fastSeedance 2 Fast480p, 720p720pPassed to the model as generate_audio
seedance_v15Seedance 1.5 Pro720p, 1080p720pToggle: true picks the audio variant
kling30_standardKling 3.0 Standard— (none)—Toggle: true picks the audio variant
kling30_proKling 3.0 Pro— (none)—Toggle: true picks the audio variant
cinematic_videoCinematic Video— (none)——
veo31_fastVeo 3.1 Fast720p, 1080p, 4k720pToggle: true picks the audio variant
veo31_liteVeo 3.1 Lite720p, 1080p720pToggle: true picks the audio variant
infinitetalkInfiniteTalk480p, 720p720p—
scail2Scail 2480p, 720p720p—
happyhorseHappyHorse720p, 1080p720p—
ray32Ray 3.2540p, 720p, 1080p720p—
gemini_omni_11_flashGemini Omni 1.1 Flash— (none)——

The catalog lists the variants offered to the CLI and agents. The API also accepts the families behind the variant ids in the table below; for example, seedance2_4k is seedance2 + 4k. An unknown video_model returns a 422 that lists every accepted family, and an unsupported resolution returns a 422 such as video_resolution must be one of 480p, 720p, 1080p for seedance25.

CLI and MCP​

The astria CLI 1.20.0 and later take the family and resolution separately:

astria video --video-model seedance2_fast --video-resolution 720p \
--video-prompt "<faceid:1234:1> woman walks down a runway" --duration 5 --wait

The Astria MCP generate_video tool accepts video_resolution alongside video_model, its generate_audio flag picks the audio variant of Toggle families, and list_models returns video_model_families. Pass num_images 1 when calling generate_video from the MCP or CLI; otherwise the image-stage default number of images is used.

Variant ids​

Each variant id is a family + resolution (+ audio) combination and remains a valid video_model. Cost is the per-prompt charge in cents at the base duration listed for the model; videos longer than the base scale linearly. _audio variants generate an audio track.

video_modelbase cost (¢)video_duration
seedance_480p102–12
seedance_v15_720p144–12
seedance_v15_audio_720p294–12
seedance2_fast_480p554–15
seedance2_fast_720p1104–15
seedance2_fast_1080p2754–15
seedance2_fast_4k5504–15
seedance2_480p664–15
seedance2_720p1324–15
seedance2_1080p3304–15
seedance2_4k6604–15
wan22_720p435
wan22_fast_480p65
wan22_fast_580p85
wan22_fast_720p115
wan25_720p535, 10
wan26_720p535, 10, 15
wan26_1080p795, 10, 15
wan27_720p555, 10, 15
wan27_1080p835, 10, 15
wan_animate_720p4410
ltx23_720p175, 10, 15, 20
ltx23_1080p225, 10, 15, 20
kling25395, 10
kling30_standard923–15
kling30_standard_audio1393–15
kling30_pro1233–15
kling30_pro_audio1853–15
kling30_4k2633–15
kling30_motion_control27710
kling30_motion_control_pro37010
cinematic_video845, 10, 15
dreamactor_m22910
happyhorse_720p773–10
happyhorse_1080p1323–10
happyhorse_motion_control15410
veo31_fast_720p854, 6, 8
veo31_fast_audio_720p1264, 6, 8
veo31_fast_1080p854, 6, 8
veo31_fast_audio_1080p1264, 6, 8
veo31_fast_4k2648
veo31_fast_audio_4k3088
veo31_lite_720p444, 6, 8
veo31_lite_audio_720p444, 6, 8
veo31_lite_1080p714, 6, 8
veo31_lite_audio_1080p714, 6, 8

Capability matrix​

  • Text-to-video (no first frame required): seedance2_*, happyhorse_720p, happyhorse_1080p.
  • Multi-reference images (image_references[] / image_reference_urls[]): seedance2_*, seedance25*, minimax_h3_2k, wan30_*, happyhorse_720p, happyhorse_1080p, ray32_*, gemini_omni_flash. Limits vary by model (nine for Seedance 2 and MiniMax H3, ten for WAN 3.0, and up to thirty for Seedance 2.5).
  • Reference video (video_references[]): seedance25*, seedance2_* and motion-control models. Several per request on seedance25 and seedance25_draft.
  • Reference audio (audio_references[]): seedance25* and seedance2_*. Several per request on seedance25 and seedance25_draft.
  • Generated audio (video_audio=true): seedance2_*.
  • First+last keyframe (video_last_frame): seedance_v15_*, seedance2_*, wan21_*, wan26_*, wan27_*, wan_fast_*, ltx23_*, kling*, veo31_*, hailuo*.
  • Motion control (requires one video_references[] item): kling30_motion_control*, wan_animate_720p, dreamactor_m2, happyhorse_motion_control.

Backwards compatibility​

The pre-2026-04 syntax embedding flags inside text is still accepted and promoted to the new columns server-side:

<lora:1533312:1.0> ohwx woman hiking in the alps --video --video_model seedance_v15_720p --duration 5 --video_prompt "Woman looking at the camera, smiling, confident"

The text syntax also accepts --video_resolution, for example --video_model seedance_v15 --video_resolution 1080p.

New integrations should prefer the dedicated form fields.