Veo/Kling/Sora Video Generation

Generate high-quality videos using:

• Veo 3.1 (Google) — 1080p videos with native audio/dialogue/music
• Kling (Kuaishou) — 1080p videos with native audio in 5 languages, 5-10s, best for long-form via extension (default model: kling-v3-0)
• Sora 2 (OpenAI) — Variable duration videos (fallback)

Preferred backend order: veo-fast (primary) → kling-pro (secondary) → sora (tertiary)

Default backend: vertex-veo-fast (Veo 3.1 Fast on Vertex AI, auto-falls back to AI Studio)

NEVER call _generate_video_internal.py or _extend_video_internal.py directly — they WILL reject the call. Use generate_video_safe.sh and extend_video_safe.sh instead.

Quick Start

SLACK_CHANNEL_ID="<channel_id>" \
SLACK_THREAD_TS="<thread_ts>" \
  ~/.openclaw-rawdog/skills/veo-video-gen/scripts/generate_video_safe.sh \
  --prompt "your video description" \
  --filename "/Users/novarussell/.openclaw-rawdog/media/$(date +%Y-%m-%d-%H-%M-%S)-video.mp4"

All Parameters

Parameter	Short	Required	Values	Default	Description
--prompt	-p	YES	text	—	Full video description including dialogue/voiceover
--filename	-f	YES	path	—	Output path (absolute, to ~/.openclaw-rawdog/media/)
--backend	-b	no	see below	vertex-veo-fast	Generation backend
--duration	-d	no	integer (seconds)	8	Video length
--aspect-ratio	-a	no	9:16, 16:9, 1:1	9:16	Aspect ratio
--resolution	-r	no	720p, 1080p	1080p	Output resolution
--num-videos	-n	no	1, 2, 3, 4	1	Number of videos to generate in one call
--person-generation	—	no	allow_all, allow_adult, dont_allow	unset	Person generation policy
--seed	—	no	integer	unset	Seed for reproducibility
--reference-image	-I	no	path	—	Reference image for image-to-video
--negative-prompt	—	no	text	—	Negative prompt (Kling V1 only, skipped on v2.5+)
--kling-model	—	no	model name	kling-v3-0	Kling model override (kling-v1, kling-v2-6, kling-video-o1, kling-v3-0)
--cfg-scale	—	no	float 0-1	0.5	CFG scale (Kling V1 only, ignored on v2+)
--camera	—	no	preset or JSON	—	Camera control for Kling text2video (V1 only, see presets below)
--multi-shot	—	no	JSON array	—	Multi-shot scenes for Kling (see below)
--end-image	—	no	path	—	End-frame image for Kling image-to-video (image_tail)
--no-upload	—	no	flag	—	Skip automatic Slack upload

Backends

Backend	Model	Route	Notes
vertex-veo-fast	veo-3.1-fast-generate-001	Vertex AI → AI Studio fallback	Default. GA model, 50 RPM
vertex-veo-quality	veo-3.1-generate-001	Vertex AI → AI Studio fallback	Higher quality, slower
veo-fast	veo-3.1-fast-generate-preview	AI Studio → Vertex fallback	AI Studio first
veo-quality	veo-3.1-generate-preview	AI Studio → Vertex fallback	AI Studio first
kling-pro	kling-v3-0 (pro mode)	Kling REST API	1080p, ~60s gen, best quality. Override with --kling-model
kling-std	kling-v3-0 (std mode)	Kling REST API	720p, ~30s gen, faster. Override with --kling-model
sora	Sora 2	OpenAI via litellm	Fallback option

Auto-fallback: vertex backends fall back to AI Studio on failure; AI Studio backends fall back to Vertex on rate limit.

Video Extension

Extend/continue a previously generated video using extend_video_safe.sh. Reads the .meta.json sidecar (auto-created by generate_video_safe.sh) to detect the backend and call the appropriate extension API.

Extension Command

SLACK_CHANNEL_ID="<channel_id>" \
SLACK_THREAD_TS="<thread_ts>" \
  ~/.openclaw-rawdog/skills/veo-video-gen/scripts/extend_video_safe.sh \
  --video "/Users/novarussell/.openclaw-rawdog/media/original-video.mp4" \
  --prompt "continuation description" \
  --extensions 3

Extension Parameters

Parameter	Short	Required	Default	Description
--video	-v	YES	—	Path to video (must have .meta.json sidecar)
--prompt	-p	no	original prompt	Continuation prompt
--negative-prompt	—	no	—	Negative prompt (Kling only)
--extensions	-n	no	1	Number of extensions to chain
--filename	-f	no	{original}_extended.mp4	Output path
--no-upload	—	no	—	Skip Slack upload

Extension Behavior Per Backend

Backend	Method	Per Extension	Max Total	Notes
Kling	Dedicated /video-extend endpoint	+5s	~3 min	Best for long-form, chains via video_id
Veo 3.1	SDK re-call with video ref	+7s	~148s (2.5 min)	720p only, requires AI Studio backend, 2-day URI expiry
Sora 2	Frame extract + image-to-video + concat	configurable	~60s practical	Coherence degrades, no audio continuity

Extension Notes

• Prerequisite: The video must have a .meta.json sidecar (auto-created by generate_video_safe.sh)
• Veo: Only AI Studio backends (veo-fast, veo-quality) save the video URI needed for extension. Vertex-generated videos cannot be extended — regenerate with an AI Studio backend first. URIs expire after 2 days.
• Sora: Uses frame extraction + image-to-video + ffmpeg concat. Requires ffmpeg installed. Calls OpenAI directly (not litellm).
• Kling: Best for long-form — dedicated /video-extend endpoint, chains via video_id (saved in .meta.json).

Multi-Video Output (--num-videos)

Generate 1–4 variations in one call:

SLACK_CHANNEL_ID="<channel>" SLACK_THREAD_TS="<thread>" \
  ~/.openclaw-rawdog/skills/veo-video-gen/scripts/generate_video_safe.sh \
  --prompt "A serene beach at sunset with gentle waves" \
  --filename "/Users/novarussell/.openclaw-rawdog/media/$(date +%Y-%m-%d-%H-%M-%S)-beach.mp4" \
  --num-videos 4 \
  --duration 8

When --num-videos > 1, files are saved as {stem}_1.mp4, {stem}_2.mp4, etc. All are uploaded to Slack individually.

Person Generation

Controls whether and what age of people can appear:

--person-generation allow_all    # All ages allowed (good for family/lifestyle content)
--person-generation allow_adult  # Adults only
--person-generation dont_allow   # No people (product shots, landscapes)

Default is unset (model decides). For ad content featuring people, use allow_adult or allow_all.

Resolution

--resolution 1080p   # Default — full HD
--resolution 720p    # Faster generation, smaller files

Seed (Reproducibility)

--seed 42   # Reproduce a specific generation
--seed 0    # Fixed seed from the docs example

Useful for iterating on a prompt while keeping similar composition.

Image-to-Video

SLACK_CHANNEL_ID="<channel>" SLACK_THREAD_TS="<thread>" \
  ~/.openclaw-rawdog/skills/veo-video-gen/scripts/generate_video_safe.sh \
  --prompt "A woman relaxing in white bedding, camera slowly dollies forward, morning sunlight streaming through windows" \
  --reference-image "/Users/novarussell/.openclaw-rawdog/media/remixed-frame.png" \
  --filename "/Users/novarussell/.openclaw-rawdog/media/$(date +%Y-%m-%d-%H-%M-%S)-from-image.mp4" \
  --duration 8 \
  --aspect-ratio 9:16 \
  --person-generation allow_adult

Kling Camera Control (--camera)

Control camera movement in Kling text-to-video generation. V1 models only (--kling-model kling-v1) — camera_control is not supported on v2+ models and will be ignored with a warning.

Named Presets (type-based, no config)

Preset	API type	Effect
down-back	down_back	Camera moves down and back
forward-up	forward_up	Camera moves forward and up
right-turn	right_turn_forward	Camera turns right moving forward
left-turn	left_turn_forward	Camera turns left moving forward

Simple Movement Presets (type="simple" with config)

Preset	Effect
dolly-in	Zoom into scene (zoom=10)
dolly-out	Zoom out of scene (zoom=-10)
pan-left	Camera pans left (pan=-10)
pan-right	Camera pans right (pan=10)
tilt-up	Camera tilts upward (tilt=10)
tilt-down	Camera tilts downward (tilt=-10)
zoom-in	Zoom into scene (zoom=10)
zoom-out	Zoom out of scene (zoom=-10)

SLACK_CHANNEL_ID="<channel>" SLACK_THREAD_TS="<thread>" \
  ~/.openclaw-rawdog/skills/veo-video-gen/scripts/generate_video_safe.sh \
  --prompt "A cat sitting on a windowsill" \
  --filename "/Users/novarussell/.openclaw-rawdog/media/cat-dolly.mp4" \
  --backend kling-pro --kling-model kling-v1 --camera dolly-in

For custom camera values, pass raw JSON matching the official API structure: --camera '{"type": "simple", "config": {"horizontal": 0, "vertical": 0, "pan": -5, "tilt": 0, "roll": 0, "zoom": 3}}'

Kling Multi-Shot (--multi-shot)

Generate multi-scene videos (3-15s total) in a single Kling call. Pass a JSON array of scenes:

SLACK_CHANNEL_ID="<channel>" SLACK_THREAD_TS="<thread>" \
  ~/.openclaw-rawdog/skills/veo-video-gen/scripts/generate_video_safe.sh \
  --filename "/Users/novarussell/.openclaw-rawdog/media/multi-shot.mp4" \
  --backend kling-pro \
  --multi-shot '[{"prompt":"Wide shot of a beach at sunrise","duration":5},{"prompt":"Close-up of waves crashing on rocks","duration":5}]'

When --multi-shot is provided, --prompt is ignored. Each shot needs prompt and duration (seconds).

Kling End-Frame (--end-image)

For Kling image-to-video, specify an end-frame image to create smooth start-to-end transitions:

SLACK_CHANNEL_ID="<channel>" SLACK_THREAD_TS="<thread>" \
  ~/.openclaw-rawdog/skills/veo-video-gen/scripts/generate_video_safe.sh \
  --prompt "Smooth zoom transition" \
  --reference-image start.png --end-image end.png \
  --filename "/Users/novarussell/.openclaw-rawdog/media/transition.mp4" \
  --backend kling-pro

Kling Model Override (--kling-model)

Default model is kling-v2-6. Override for newer/different models:

--kling-model kling-video-o1   # Reasoning model
--kling-model kling-v2-5-turbo # Turbo variant
--kling-model kling-v3-0       # V3 (if available on your account)

Audio / Voiceover

Veo 3.1 generates audio natively (dialogue, sound effects, ambient, music). Audio is always enabled on Vertex backends. Kling v2-6+ also generates native audio via enable_audio: true (sent automatically for v2+ models).

• Include spoken dialogue directly in the prompt in quotes
• Example: A woman says, "Your bedroom really says a lot about you."
• Keep spoken lines under ~7 words each for best lip-sync
• Kling: Audio is auto-enabled on v2-6+ models. Not available when using --end-image (image_tail is incompatible with enable_audio). V1 models do not support audio.

Full-Featured Example

SLACK_CHANNEL_ID="<CHANNEL_ID>" \
SLACK_THREAD_TS="<THREAD_TS>" \
  ~/.openclaw-rawdog/skills/veo-video-gen/scripts/generate_video_safe.sh \
  --prompt "The camera dollies to show a close up of a desperate man in a green trench coat making a call on a rotary wall phone with green neon light, cinematic movie scene" \
  --filename "/Users/novarussell/.openclaw-rawdog/media/$(date +%Y-%m-%d-%H-%M-%S)-trenchcoat.mp4" \
  --backend vertex-veo-quality \
  --duration 8 \
  --aspect-ratio 9:16 \
  --resolution 1080p \
  --num-videos 4 \
  --person-generation allow_all \
  --seed 0

Output

• Saves MP4(s) to specified path (or {stem}_1.mp4, {stem}_2.mp4 for multi-video)
• Saves .meta.json sidecar alongside each video (for extension support)
• Uploads each video to Slack automatically (unless --no-upload)
• Prints MEDIA:<full_path> for each video
• Prints SLACK_UPLOADED:true on successful upload

Timing

• Veo: 60–120 seconds per generation
• Kling: ~30s (std) to ~60s (pro) per generation
• Sora: 60–90 seconds
• For --num-videos 4: allow up to 10 minutes (videos generate in parallel on Vertex)
• Use a yield/timeout of at least 600000ms (10 minutes) for multi-video or extension calls

CRITICAL RULES

• NEVER modify, shorten, or simplify the user's prompt. Pass it exactly as provided.
• NEVER remove dialogue, voiceover, or narration from the prompt.
• NEVER edit this script. If it fails, try a different backend or report the error.
• If one backend fails, try another (vertex-veo-fast → vertex-veo-quality → kling-pro → sora).

API Keys / Auth

• Vertex: GOOGLE_APPLICATION_CREDENTIALS → ~/.config/gcloud/vertex-service-account.json
• AI Studio: GEMINI_API_KEY
• Kling: KLING_ACCESS_KEY + KLING_SECRET_KEY (JWT auth)
• Sora: OPENAI_API_KEY (via litellm proxy)