CogVideoX 3 is Zhipu AI’s video model with documented text-to-video, image-to-video and first-and-last-frame generation. The maker highlights frame stability, instruction following and realistic or three-dimensional visual styles. Choose text when the scene begins as a written brief, one image when its visual basis already exists, or two endpoint images when the sequence needs a defined opening and closing view.
For a text brief, describe the subject and one readable action. State the setting, viewpoint and important physical relationships so the scene can be judged against a clear intention. A character picking up an object needs a different sequence from a camera revealing that object. Describe both only when their relationship matters to the shot.
For an image-based sequence, prepare the reference so the subject’s distinguishing features are visible. The maker accepts PNG, JPEG and JPG images no larger than 5 MB, supplied through a URL or encoded image data. Keep the actual reference with the generated result so appearance and framing can be compared with the source rather than with a remembered version of it.
The maker’s endpoint-frame mode accepts two images in order: the first establishes the opening frame and the second the ending frame. Prepare them as the two intended states of one shot. Describe the transition or action that connects them, then review whether the subject remains coherent between the endpoints. A pair of good still images does not by itself show that the intervening movement fits the desired story.
The prompt can contain up to 512 characters. Use the available text for the action, its direction and the visual details that distinguish a successful result. Keep the brief focused on what should happen on screen. When the input images already establish subject appearance, describe the motion and timing that need to develop between or from those images.
The maker documents five- or ten-second output, with five seconds as the default, and 30 or 60 frames per second. Supported sizes include 1280×720, 720×1280, 1024×1024, 1920×1080, 1080×1920, 2048×1080 and 3840×2160. Choose the size and timing for the destination, then inspect the completed clip for framing and the pace of the actual action.
Quality and speed modes are documented separately from the optional AI sound-effect setting. Keep these choices with the prompt, images and result. The maker’s visual and input capabilities describe the model; the available controls of a particular generation method are supplied by that method’s contract. Preserve the settings actually used when comparing candidate shots.
Retain the original task identity until the video is complete and save the returned footage for editing. Review the whole sequence, including subject appearance during large movement, interactions with scene objects and the final composition. For endpoint-frame work, inspect both endpoints and the transition between them. Revise the specific visual mismatch while retaining the successful references and shot plan, and compare the completed videos rather than only their preview images.
", "aboutText": "CogVideoX 3 is Zhipu AI’s video model with documented text-to-video, image-to-video and first-and-last-frame generation. The maker highlights frame stability, instruction following and realistic or three-dimensional visual styles. Choose text when the scene begins as a written brief, one image when its visual basis already exists, or two endpoint images when the sequence needs a defined opening and closing view.\n\nFor a text brief, describe the subject and one readable action. State the setting, viewpoint and important physical relationships so the scene can be judged against a clear intention. A character picking up an object needs a different sequence from a camera revealing that object. Describe both only when their relationship matters to the shot.\n\nFor an image-based sequence, prepare the reference so the subject’s distinguishing features are visible. The maker accepts PNG, JPEG and JPG images no larger than 5 MB, supplied through a URL or encoded image data. Keep the actual reference with the generated result so appearance and framing can be compared with the source rather than with a remembered version of it.\n\nThe maker’s endpoint-frame mode accepts two images in order: the first establishes the opening frame and the second the ending frame. Prepare them as the two intended states of one shot. Describe the transition or action that connects them, then review whether the subject remains coherent between the endpoints. A pair of good still images does not by itself show that the intervening movement fits the desired story.\n\nThe prompt can contain up to 512 characters. Use the available text for the action, its direction and the visual details that distinguish a successful result. Keep the brief focused on what should happen on screen. When the input images already establish subject appearance, describe the motion and timing that need to develop between or from those images.\n\nThe maker documents five- or ten-second output, with five seconds as the default, and 30 or 60 frames per second. Supported sizes include 1280×720, 720×1280, 1024×1024, 1920×1080, 1080×1920, 2048×1080 and 3840×2160. Choose the size and timing for the destination, then inspect the completed clip for framing and the pace of the actual action.\n\nQuality and speed modes are documented separately from the optional AI sound-effect setting. Keep these choices with the prompt, images and result. The maker’s visual and input capabilities describe the model; the available controls of a particular generation method are supplied by that method’s contract. Preserve the settings actually used when comparing candidate shots.\n\nRetain the original task identity until the video is complete and save the returned footage for editing. Review the whole sequence, including subject appearance during large movement, interactions with scene objects and the final composition. For endpoint-frame work, inspect both endpoints and the transition between them. Revise the specific visual mismatch while retaining the successful references and shot plan, and compare the completed videos rather than only their preview images.", "editorialOverlay": { "copy_fields": null, "copy_sha256": null, "copy_source": null, "html_sha256": "bd2ab3453cd849ec8dac6c5dd9e5f33b46881ba0ffe1b4866064a5a723cc81b9", "module_sha256": "ee9dc1c50b1e6a8fb7fa062030fcb4646b9aaaeb08be513b723a5a92a7290f06", "raw_model_unchanged": true, "schema": "ailoot-native-model-editorial-v1" }, "locale": "en", "methods": [ { "html": "cogvideox-3CogVideoX 3 creates video from text, an image or a first-and-last-frame pair in the maker’s documented workflow. It supports five- or ten-second sequences, several output sizes up to 4K, 30 or 60 fps and optional generated sound effects.
This description covers the model’s capabilities. Available in Ailoot inputs, outputs and options are listed for the selected method below.
CogVideoX 3 is Zhipu AI’s video model with documented text-to-video, image-to-video and first-and-last-frame generation. The maker highlights frame stability, instruction following and realistic or three-dimensional visual styles. Choose text when the scene begins as a written brief, one image when its visual basis already exists, or two endpoint images when the sequence needs a defined opening and closing view.
For a text brief, describe the subject and one readable action. State the setting, viewpoint and important physical relationships so the scene can be judged against a clear intention. A character picking up an object needs a different sequence from a camera revealing that object. Describe both only when their relationship matters to the shot.
For an image-based sequence, prepare the reference so the subject’s distinguishing features are visible. The maker accepts PNG, JPEG and JPG images no larger than 5 MB, supplied through a URL or encoded image data. Keep the actual reference with the generated result so appearance and framing can be compared with the source rather than with a remembered version of it.
The maker’s endpoint-frame mode accepts two images in order: the first establishes the opening frame and the second the ending frame. Prepare them as the two intended states of one shot. Describe the transition or action that connects them, then review whether the subject remains coherent between the endpoints. A pair of good still images does not by itself show that the intervening movement fits the desired story.
The prompt can contain up to 512 characters. Use the available text for the action, its direction and the visual details that distinguish a successful result. Keep the brief focused on what should happen on screen. When the input images already establish subject appearance, describe the motion and timing that need to develop between or from those images.
The maker documents five- or ten-second output, with five seconds as the default, and 30 or 60 frames per second. Supported sizes include 1280×720, 720×1280, 1024×1024, 1920×1080, 1080×1920, 2048×1080 and 3840×2160. Choose the size and timing for the destination, then inspect the completed clip for framing and the pace of the actual action.
Quality and speed modes are documented separately from the optional AI sound-effect setting. Keep these choices with the prompt, images and result. The maker’s visual and input capabilities describe the model; the available controls of a particular generation method are supplied by that method’s contract. Preserve the settings actually used when comparing candidate shots.
Retain the original task identity until the video is complete and save the returned footage for editing. Review the whole sequence, including subject appearance during large movement, interactions with scene objects and the final composition. For endpoint-frame work, inspect both endpoints and the transition between them. Revise the specific visual mismatch while retaining the successful references and shot plan, and compare the completed videos rather than only their preview images.
POST /v1/videostextvideo| Field | Type | Requirement | Values & limits | Default |
|---|---|---|---|---|
Fieldmodel | Typestring | RequirementRequired | Values & limits cogvideox-3 Conditions: Use the exact Ailoot model selector. | Default— |
Fieldprompt | Typestring | RequirementRequired | Values & limits Conditions: Describe the scene and motion. | Default— |
Fieldquality | Typeboolean | RequirementOptional | Values & limits Conditions: Optional quality flag; this selected schema supplies no compatible enum or default. | Default— |
Fieldwith_audio | Typeboolean | RequirementOptional | Values & limits Conditions: Generate AI sound effects when enabled. | Defaultfalse |
Fieldsize | Typestring | RequirementOptional | Values & limits Conditions: Supply the documented size string when selecting explicit dimensions. | Default— |
Fieldfps | Typestring | RequirementOptional | Values & limits 30, 60 Conditions: Frame rate is represented as a string on this selected native method. | Default30 |
Fieldduration | Typestring | RequirementOptional | Values & limits 5, 10 Conditions: Duration in seconds is represented as a string on this selected native method. | Default5 |
| Parameter | Contract |
|---|---|
Parametercontent_type | Contractapplication/json |
Parameterrequired | Contractmodel, prompt |
Parameteroptional | Contractquality, with_audio, size, fps, duration |
Parametermaterial_types | Contracttext |
Parameterconditions | ContractUse the exact Ailoot model selector., Describe the scene and motion., Optional quality flag; this selected schema supplies no compatible enum or default., Generate AI sound effects when enabled., Supply the documented size string when selecting explicit dimensions., Frame rate is represented as a string on this selected native method., Duration in seconds is represented as a string on this selected native method. |
Method aliases: /v1/videos/generations
Lifecycle: submit_once_poll_original
GET /v1/videos/{request_id}
GET /v1/videos/generations/{request_id}
GET /v1/videos/{request_id}/content
GET /v1/media/assets
GET /v1/media/assets/{asset_id}
The displayed rates estimate the cost before generation. The final charge is calculated from the cost recorded for the original generation. Measured output duration and forecast quantities are not the final billable quantity.
Tariff references do not add request options. Supported parameters are listed for the selected method.
Match each charge to the conditions shown. Use one rate for each set of alternative configurations.
| Configuration | Rate schedule | Forecast quantity | Billing quantity | Charge |
|---|---|---|---|---|
| ConfigurationBase token rates | Rate scheduleBase rate | Forecast quantity— | Billing quantity— | Charge Input · Input tokens · Estimate for comparison only — / 1K tokens Output · Output tokens · Estimate for comparison only — / 1K tokens Input · Cache read · Estimate for comparison only — / 1K tokens Input · Cache write · Estimate for comparison only — / 1K tokens Input · Cache write, 1 hour · Estimate for comparison only — / 1K tokens |
| ConfigurationPer request | Rate schedule— | Forecast quantity— | Billing quantityone request | Charge Input · Estimate for comparison only $0.3 per request |
| ConfigurationEstimated generation cost | Rate schedule— | Forecast quantityone generation estimate | Billing quantity— | Charge Output · Estimated generation cost · Estimate for comparison only $0.3 estimate before generation This estimate is not an additional charge or a fixed rate for the completed generation. |
Ailoot reserves an estimated amount before a request. After the result, it settles the final charge from measured usage. Streaming settlement uses the final usage record. How reservation and settlement work →
set -eu\n: "${AILOOT_API_KEY:?Set your Ailoot API key}"\ncat > request.json <<'JSON'\n{\n "model": "cogvideox-3",\n "prompt": "A cat is walking on grass."\n}\nJSON\ncurl --fail-with-body --silent --show-error https://api.ailoot.ai/v1/videos \\\n -H "Authorization: Bearer $AILOOT_API_KEY" -H 'Content-Type: application/json' \\\n --data-binary @request.json -D response-headers.txt -o result.json\n# Create once; retain id, polling_url, original task identifiers and every header.\nset -eu\n: "${AILOOT_API_KEY:?Set your Ailoot API key}"\nexport AILOOT_API_KEY\npython3 - <<'PY_STATUS'\nimport json,os,subprocess\nfrom pathlib import Path\nfrom urllib.parse import urljoin,urlsplit\nr=json.loads(Path("result.json").read_text())\nu=urljoin("https://api.ailoot.ai",r["polling_url"])\np=urlsplit(u)\nif p.scheme!="https" or p.netloc!="api.ailoot.ai" or p.username or p.password:\n raise SystemExit("Unexpected returned polling URL; preserve the response")\nsubprocess.run(["curl","--fail-with-body","--silent","--show-error",u,"-H","Authorization: Bearer "+os.environ["AILOOT_API_KEY"],"-D","status-headers.txt","-o","status.json"],check=True)\n# Read this same original polling_url again while pending.\n# Preserve unknown/result_pending_check; do not repeat creation.\nPY_STATUS\nset -eu\n: "${AILOOT_API_KEY:?Set your Ailoot API key}"\nexport AILOOT_API_KEY\npython3 - <<'PY_ASSETS'\nimport json,os,subprocess\nfrom pathlib import Path\nfrom urllib.parse import urljoin,urlsplit\nr=json.loads(Path("status.json").read_text())\nroles={}\ndef visit(v,path=""):\n if isinstance(v,dict):\n for k,x in v.items(): visit(x,path+"/"+k)\n elif isinstance(v,list):\n for i,x in enumerate(v): visit(x,path+"/"+str(i))\n elif isinstance(v,str):\n u=urljoin("https://api.ailoot.ai",v); p=urlsplit(u)\n if p.scheme=="https" and p.netloc=="api.ailoot.ai" and not p.username and not p.password and p.path.startswith("/v1/media/assets/") and p.path.endswith("/content"):\n roles.setdefault(u,[]).append(path)\nvisit(r)\nif not roles: raise SystemExit("No completed owned assets; retain the original status")\nmanifest=[]\nfor i,(u,paths) in enumerate(roles.items()):\n target=Path(f"asset-{i}.bin")\n subprocess.run(["curl","--fail-with-body","--silent","--show-error",u,"-H","Authorization: Bearer "+os.environ["AILOOT_API_KEY"],"-o",str(target)],check=True)\n raw=target.read_bytes()\n if not raw: raise SystemExit("Empty asset; retain original status and metadata")\n ext="glb" if raw[:4]==b"glTF" else "png" if raw[:8]==b"\\x89PNG\\r\\n\\x1a\\n" else "jpg" if raw[:3]==b"\\xff\\xd8\\xff" else "webp" if raw[:4]==b"RIFF" and raw[8:12]==b"WEBP" else "mp4" if raw[4:8]==b"ftyp" else "bin"\n name=f"asset-{i}.{ext}"\n if target.name!=name: target.rename(name)\n manifest.append({"file":name,"response_paths":paths,"bytes":len(raw)})\nPath("asset-manifest.json").write_text(json.dumps(manifest,indent=2))\nprint("Saved",len(manifest),"unique owned assets; every response role retained")\nPY_ASSETS\nGET /v1/videos/{request_id}GET /v1/videos/generations/{request_id}GET /v1/videos/{request_id}/contentGET /v1/media/assetsGET /v1/media/assets/{asset_id}Save every completed video or returned file; retrieve every actual returned asset URL.
Retain the complete response and headers, original identifiers, every actual output and all usage/accounting fields. Save each unique owned asset and retain every role pointing to it.
Keep the returned public task or operation id and all headers. Read only the original task; unknown or result_pending_check does not authorize another creation.
| Field | Type |
|---|---|
Fieldid | Typestring |
Fieldpolling_url | Typestring |
Fieldprovider_task_id | Typestring |
Fieldid | Typestring |
Fieldtask_status | Typestring |
Fieldvideo_result[].url | Typestring |
Ailoot Ultra cogvideox-3
/v1/videos/generations