Mode d’emploi
- Copiez la commande et examinez le paquet avant de l’installer dans votre environnement agent.
- Consultez le document original pour les outils, identifiants et environnements requis.
- Décrivez les entrées et le livrable souhaité ; comparez la sortie aux sources.
npx skills add https://github.com/openmontageinc/openmontage --skill pika-h3-flexAvant de commencer
- L’installation n’inclut ni abonnement au modèle, ni quota API, ni accès aux fichiers.
- Étoiles et installations reflètent l’activité, pas un test indépendant de qualité ou de sécurité.
Rendu depuis le dépôt source en conservant titres, exemples, code, tableaux, liens et images.
Pika H3 Flex
Pika H3 Flex is Pika's post-trained H3 model for professional iteration: six sampling steps, synchronized stereo audio (dialogue, ambience, music) in one pass, and shot direction through references and prompts. pika_video calls the hosted Pika API; the open weights are also published on Hugging Face for local ComfyUI use, which is a separate, unverified path.
Route and billing
| Item | Value |
|---|---|
| Tool | pika_video, model: "pika-h3-flex" |
| Credential | PIKA_API_KEY from https://dev.pika.art (API Keys) |
| Support | support@pika.art for API, billing and key problems |
| Billing | Prepaid per output second, charged to the user's own Pika organization. Failed jobs are not charged. A job covered by a Pika promotional allowance settles at $0 and reports data.free_generation. |
| Price | 480p $0.01/s, 768p $0.02/s, 1080p $0.035/s, 2k $0.04/s, 4k $0.08/s. Reference-to-video: 480p $0.03/s, 768p $0.06/s, 1080p $0.075/s, 2k $0.08/s, 4k $0.10/s, plus reference video input seconds at the same rates. |
| Cost reporting | cost_usd is the settled charge from the job record. data.cost_status is settled, pending or unavailable; only settled is final. |
OpenMontage's default budget policy asks the user before any single action over $0.50 and the first time a new paid tool is used. A 768p 15 second clip is $0.30 and runs unprompted; any 4k clip of 7 seconds or more prompts.
Set dry_run: true to validate the request and get the exact quoted price from Pika's free quote endpoint without uploading or generating. For reference-to-video with video references, pass reference_video_seconds so the preflight estimate includes the input-video charge.
Operations
| Operation | Inputs | Notes |
|---|---|---|
text_to_video | prompt | Default when nothing else is supplied. |
image_to_video | prompt + image_url or image_path (first frame), or last_frame_image_url / last_frame_image_path | Either frame alone is valid. reference_image_url / reference_image_path are accepted as first-frame aliases. |
first_last_frame_to_video | prompt + both frames | Describe the motion between the two frames. |
reference_to_video | prompt + references | references in Pika's native ordered shape, or reference_image_urls / reference_video_urls / reference_audio_urls and the *_paths equivalents. Local files are uploaded to Pika. |
Operation is inferred from the inputs when omitted. Supplying frames with operation: "text_to_video" is an error, not a silent downgrade.
Reference rules
- At most 9 images, 3 videos and 3 standalone audio clips, 12 files total.
- At least one image or video is required; audio alone is rejected locally.
- Each video or audio clip must be 2 to 15 seconds; combined video and combined
standalone audio must each fit in 15 seconds. A video's own soundtrack stays with that video and is not counted as a standalone audio reference.
- Order matters. The per-type lists are sent as images, then videos, then audio.
Use references when a different order is intended.
Shared parameters
duration: whole seconds, 3 to 15, default 5. Output rounds up to the
model's frame grid (124 frames for a 5 second request at 24 fps), so clips run slightly long. Plan edits on the delivered length, not the requested one.
resolution:480por768pnative;1080p,2k,4kare generated at
768p and upscaled once, preserving audio and frame rate. Draft at 480p, then rerun the approved take at the delivery resolution.
aspect_ratio:auto,21:9,16:9,3:2,4:3,5:4,1:1,4:5,
3:4, 2:3, 9:16, 9:21. Omitted: auto when rewriting, otherwise the first keyframe's ratio or 16:9. Keyframes are resized and center-cropped.
seed: integer. Same seed and same request reproduce closely, not exactly.rewrite: defaulttrue. Pika improves the prompt before generation. Turn it
off for exact prompt control, for A/B tests across seeds, or when the prompt already follows the structure below. prompt_optimizer is accepted as an alias.
- Audio is always generated.
generate_audio: falseis rejected; strip audio in
post with video_compose if a silent clip is needed.
Prompting
Write the prompt as a shot list, not a caption. The model follows explicit structure well:
- Scene: setting, time of day, lighting, lens or style in one sentence.
- Subjects: who or what, with the one or two details that must hold.
- Action and camera: what happens, in order, with camera moves named
(slow push-in, orbit, handheld follow, whip pan, hard cut to).
- Audio: dialogue in quotes with the speaker named, then ambience and music
intent. Dialogue is lip-synced when the speaker is on screen.
- Text on screen: exact words, placement, and when they appear, for motion
graphics.
Multi-shot stories work in one prompt: separate shots with "Cut to:" and keep each shot's subject and lighting consistent by repeating the locks. For reference-guided work, say what each reference is for ("first image: the product; video: the dance to follow; audio: the track to cut to").
Keep prompts under 8,000 characters. Describe motion that fits the duration; a 15 second request is the ceiling, not a target.
Failure handling
insufficient_balanceorcycle_limit_exceeded: the user's Pika balance is
empty. Point them to https://dev.pika.art. Do not switch provider silently.
rate_limited(HTTP 429): honordata.retry_after, then submit again.content_moderationorinvalid_input: revise the prompt or inputs; the
message names the problem.
- Poll timeout: the result carries
data.resume_job. Callpika_videoagain
with that object (and a longer poll_timeout) to keep waiting on the same job instead of paying for a new one. Set job_path to persist the checkpoint.
- The tool never retries a paid submission on its own.
Provenance
Every result records data.request (the exact payload sent), data.job_id, data.usage, data.billing and data.resume_job. Keep these with the clip for reproducibility and cost reconciliation.