VORQ DocsGenerate images and video
Submit image and video jobs, attach reference assets and save the returned frames.
Media jobs use the same client.submit() call as text. The difference is the input: pass a
dict in the model's own input shape, and the client reads that shape to decide what the order
buys. Pick a model id from await client.models.list().
Generate an image
handle = await client.submit(
model=IMAGE_MODEL, # an image model id from client.models.list()
input={
"prompt": "a lighthouse in fog, oil painting",
"width": 1024,
"height": 768,
"num_images": 1,
"seed": 42,
},
sla="batch",
rate_out=50_000, # atomic token units per 1,000,000 output pixels
)
result = await handle.result() # a MediaResult
paths = result.download("out/") # writes out/0.png, ...
print(paths, result.seed, result.cost)An image order buys width × height × num_images output pixels, so rate_out is a price per
million pixels. Width and height each default to 1024 when you leave them out.
Generate a video
A request becomes a video order when it names duration (or duration_secs):
handle = await client.submit(
model=VIDEO_MODEL,
input={
"prompt": "a slow pan across a mountain valley at dawn",
"resolution": "720p",
"aspect_ratio": "16:9",
"duration": 5,
},
sla="batch",
rate_out=2_000, # atomic token units per 1,000,000 pixel-seconds
)
result = await handle.result()
frame = result.frames[0] # a video comes back as a single frame entry
print(frame["content_type"], frame["duration_secs"])
result.download("out/")A video order buys frame pixels × seconds. Here that is 1280 × 720 × 5. The resolution
tiers, aspect ratios and defaults are listed in Units and media inputs.
Attach reference assets
Reference images, clips and audio travel as base64 bytes inside the sealed payload, never as URLs. Each asset declares its own media type and dimensions:
import base64
from pathlib import Path
def b64(path: str) -> str:
return base64.b64encode(Path(path).read_bytes()).decode()
handle = await client.submit(
model=VIDEO_MODEL,
input={
"prompt": "the flower opens",
"image": {"b64": b64("first.png"), "media_type": "image/png", "width": 1280, "height": 720},
"end_image": {"b64": b64("last.png"), "media_type": "image/png", "width": 1280, "height": 720},
"resolution": "720p",
"duration": 5,
},
rate_in=1_000, # atomic token units per 1,000,000 reference pixel-seconds
rate_out=2_000,
)References are the input side of a media order: each still counts its pixels once, each clip its
pixels times its seconds. A clip must state duration_secs. The provider decodes every asset
after opening the payload and fails the job if an asset is larger or longer than declared.
Asset keys and limits are in Units and media inputs.
Read the frames
The frames are inside the sealed result, so reading them makes no network call:
result.frames: one dict per image (or one for a video) withb64,content_type,widthandheight, plusduration_secsfor video.result.bytes(): the decoded frames, in order.result.download(dir): writes each frame as0,1, … with the file suffix itscontent_typeimplies, and returns the paths.
See MediaResult.