guaardvark / video

Video with Guaardvark

Read setup first if the backend or the comfyui plugin state is unknown.
Video needs a 16 GB-class NVIDIA card. Clips take minutes, so every route is queued.

Pick the model

GET ${GUAARDVARK_URL:-http://localhost:5000}/api/batch-video/models lists the registry with
is_downloaded / is_ready (and missing_files), capabilities (modes t2v/i2v, max_frames, native_fps,
aspect_ratios, min_steps, speed_profiles, audio_out), vram_mb, size_gb, license. Installed on a typical box:

The active default is GET /api/settings/active_video_model (resolved.t2v, resolved.i2v, resolved.scene).
Never pass a step count below the model's min_steps; the server raises it and the result would be smeared anyway.

One clip: MCP generate_video

• prompt: scene, subject, motion, style, and any spoken lines (H3 speaks them).

• model, duration_s (clamped to the model), aspect_ratio (one the model declares),
  style (cinematic, realistic, anime, 3danimation, ...), `numinferencesteps,speedprofile`.

• audio: true forces a soundtrack-capable model (H3) and fails on a silent family.
• first_image / last_image: document id or path; last frame needs a first+last mode model.

• reference_images (up to 9), reference_clips (up to 3, 2-15 s, their own sound left out) and
  reference_audio (up to 3): lock a person, look, motion or voice on the reference build
  (minimax-h3-ref2va-int8, picked when no model is named). The prompt names them
  <Picture N>, <Video N>, <Audio N>; to keep a clip's sound, pass its track as
  reference_audio. In the Studio this is the References input on Video Gen.

• wait_for_result default false: the tool returns a batch id and a Studio deep link at once.
  Give the user the link; poll get_generation_status(batch_id=...) (MCP) or
  GET /api/batch-video/status/<batch_id> if they ask you to wait. wait_for_result: true
  blocks for the clip, up to 30 minutes.

Looping animation: MCP generate_animation

Frame-morph GIF/MP4 via img2img: prompt, motion, frames 2-24, strength 0.1-0.5, format gif|mp4|both.
Use for short loops and stickers, not for cinema clips. Over MCP it answers at once with a job id
(tooljob_...); poll get_generation_status for the GIF and MP4 URLs, or pass
wait_for_result: true to wait up to 60 s for them.

Batch: REST

B=${GUAARDVARK_URL:-http://localhost:5000}
# text to video, one clip per prompt
curl -s -X POST $B/api/batch-video/generate/text -H 'Content-Type: application/json' -d '{
  "prompts": ["a red kite over a grey sea", "the same kite at dusk"],
  "model": "wan22-5b", "prompt_style": "cinematic", "enhance_prompt": true, "seed": 42
}'
# image to video
curl -s -X POST $B/api/batch-video/generate/image -H 'Content-Type: application/json' -d '{
  "image_paths": ["/abs/path/frame.png"], "prompt": "slow push in, wind in the grass", "model": "wan22-5b"
}'

Optional keys the server honours: negative_prompt, guidance_scale, motion_strength,
interpolation_multiplier (RIFE frame interpolation), combine_frames, lora_name +
lora_strength, adapters, guides (per-prompt audio/image anchors on models that declare them),
last_frame_paths (image route), storyboard_concept / storyboard_shots.
Poll GET $B/api/batch-video/status/<batch_id>; files via
GET $B/api/batch-video/video/<batch_id>/<video_name>; cancel POST $B/api/batch-video/batch/<batch_id>/cancel;
retry POST $B/api/batch-video/retry/<batch_id>.

Rules

• Quote the model and the queued batch id back to the user. Never claim a clip is done until the
  status route says so.

• The GPU is exclusive: while video renders, chat models are evicted. Warn before queuing a long batch.
• The Studio page (Video Gen) shows the same queue; the user may prefer to watch it there.