AI video generation: text-to-video, image-to-video, video-to-video, model selection. Use when generating a short video clip from a prompt or reference (e.g. 5s clip of a cat in rain, animate this photo, restyle this video).

video Use this skill for all video generation requests on Starchild. Core principle: call the provided scripts. Do not re implement proxy/billing/upload plumbing. 1. Text to video (most common) ⚠️ Execution context — read this first. The code blocks below are Python , not shell commands. Starchild's bash tool runs /bin/bash c , which cannot parse exec(open(...)) — pasting them directly into a bash command will fail with syntax error near unexpected token 'open' . Also, exec(open(...)) inside python3 c fails with NameError: file because the script uses file for path resolution. Use python3 <<'EOF' with from exports import when calling via the bash tool: The heredoc ( <<'EOF' ) preserves all quotes and newlines — no escaping needed. Note: video skill has no exports.py — import directly from generate video . generate video automatically: submits → polls → fetches result → downloads mp4 to output/videos/ . Delivering the result to the user — IMPORTANT Never hand the user the raw video url (e.g. https:// .fal.media/.../ .mp4 ). fal serves these files with Content Security Policy: sandbox; default src 'none' , which means: Opening the link in a browser shows a blank page (no inline player triggered). Embedding via <video / <iframe is blocked by CSP. There is no Content Disposition: attachment header, so the browser does not auto download either. URL side tweaks (query params, ?download=1 , etc.) cannot fix this — only a server side header change would, and we don't control fal's CDN. The only reliable user facing delivery path is the already downloaded local file : 1. Use result["local path"] (e.g. output/videos/xxx.mp4 ) — generate video always downloads on success. 2. Tell the user the file is saved to output/videos/<filename and is viewable in the workspace file panel / file browser. 3. On Web channel, also embed it inline so the user can preview it in chat: (or link as [video](output/videos/<filename .mp4) — the workspace serves these directly with the right headers). 4. On Telegram / WeChat: send the file via send to telegram(file path="output/videos/...", message type="video") or send to wechat(file path="output/videos/...", message type="video") . If the download somehow failed ( local path missing) — re fetch with: Then deliver the local path. Still do not give the user the raw fal URL as the primary deliverable. 2. Image to video / video to video (reference assets) fal.ai needs the reference asset as a public https URL . fal storage upload requires a Serverless permission your key currently does not have. The reliable path is to expose the asset via a published Starchild preview . Standard procedure 1. Drop or copy the asset into output/fal assets/ using publish asset.py . 2. Make sure a preview named fal assets is running and published (one time setup, see §3). 3. Build the public URL as <preview base /<filename . 4. Call generate video(... image url=public url) . generate video auto rewrites the model path from /text to video to /image to video whenever image url is provided. The same approach works for video to video models — pass an mp4 URL instead. Asset constraints (enforced by publish asset.py ) Image: .jpg .jpeg .png .webp .gif .bmp , max 10 MB Video: .mp4 .mov .webm .mkv .m4v , max 100 MB Anything outside these is rejected before publish 3. One time fal assets public preview setup Run this once per workspace. The preview keeps running across sessions. After publish, the public base URL is reusable for every future image to video / video to video task. Files dropped into output/fal assets/ become reachable as <base /<filename immediately — no re publish needed. Verify with: If preview(action='serve') returns No available ports in pool , ask the user which existing preview can be stopped to free a port — never silently kill one. 4. Model selection Tier Model Cost / 5s Notes budget fal ai/wan/v2.5/text to video $0.25 Fastest, cheapest; good for prompt iteration balanced alibaba/happy horse/text to video $0.70 Default; best lip sync, most use cases premium bytedance/seedance 2.0/fast/text to video $1.20 Best motion + camera direction mini bytedance/seedance 2.0/mini/text to video $0.36 (480p) / $0.77 (720p) Cheapest Seedance; resolution tiered, no 1080p. Duration must be a string ( "5" , not 5 or "5s" ) — see gotcha below premium 25 bytedance/seedance 2.5/text to video token based Supports text to video, image to video, and reference to video. Requires resolution ( 480p / 720p ), aspect ratio (six supported ratios), and integer duration from 4–30 seconds. Estimate with estimate cost(..., aspect ratio=...) . — xai/grok imagine video/v1.5/image to video $0.41 (480p) / $0.71 (720p) per 5s image to video ONLY (single required image url , no image urls ); +$0.01 input image surcharge included in estimate. ⚠️ resolution="1080p" is schema valid upstream but has NO published price — the proxy rejects it 400 fail closed — fal ai/kling video/v3/turbo/standard/text to video $0.56 per 5s Kling v3 Turbo Standard, flat $0.112/s; .../turbo/pro/... = $0.14/s ($0.70/5s); .../v3/4k/... = $0.42/s ($2.10/5s). i2v variants exist for all — alibaba/happy horse/v1.1/text to video $0.70 (720p) / $0.90 (1080p) per 5s v1.1 has its own 1080p tier $0.18/s (NOT the v1.0 2× rule); also /image to video , /reference to video — fal ai/minimax h3/text to video proxy pricing applies Supports text to video, image to video, and reference to video. For reference to video pass image urls=[...] ; the payload is translated to upstream reference image urls . ⚠️ Happy Horse default resolution is 1080p upstream (v1.0 and v1.1): omitting resolution bills the 1080p tier (v1.1 5s = $0.90; v1.0 ref2v 5s = $1.40). Pass resolution="720p" explicitly for the cheaper rate. Invalid resolution values are rejected 400 by the proxy. Reference to video : pass image urls=[...] (list of 1–9 public HTTP(S) URLs) — NOT the single image url param. generate video() validates count and URL scheme. Happy Horse and Seedance 2.5 submit the image urls field; MiniMax H3 ( fal ai/minimax h3/reference to video ) submits upstream's reference image urls field. Seedance 2.5 example : Use an integer duration from 4–30 seconds. resolution must be 480p or 720p ; aspect ratio must be one of 21:9 , 16:9 , 4:3 , 1:1 , 3:4 , 9:16 . The proxy rejects auto values because they cannot be priced safely. Override by passing the full model id to generate video(model=...) . Image to video variants are auto derived by replacing text to video with image to video . Pricing details and model registry live in generate video.py::estimate cost . For models not yet registered there, the legacy fallback is only a rough estimate and may differ from the proxy; do not use it for budgeting new endpoints. 5. Polling an existing request Use this when an earlier generate video call timed out or you only have a request id . 6. Provided scripts generate video.py — submit → poll → download. Handles text to video and image to video. publish asset.py — copy local files (or download remote URLs) into output/fal assets/ so they can be served by the fal assets preview. poll status.py — resume polling by request id , downloads the result on completion. 7. Troubleshooting Problem Fix image url must be a public HTTP(S) URL Use publish asset.py + fal assets preview, then pass the public URL No available ports in pool (preview serve) Ask the user which preview to stop; do not auto kill downstream service error after COMPLETED Reference asset host failed mid render — re encode/resize to 16:9, re publish, retry HTTP 402 insufficient credits Top up balance; cost is pre charged on submit HTTP 403 endpoint not allowed sc proxy only allows approved fal video endpoints; pick one from the model table Generation FAILED upstream Shorten prompt, drop unusual tokens, retry once before changing model HTTP 422 literal error on duration (Seedance Mini) Mini requires duration as a string ( "5" , "10" , "auto" ), not an int and not "5s" . generate video() encodes this automatically when model contains seedance 2.0/mini — only hit this if you hand build the request body. Other Seedance variants accept int/ "5s" as before. Seedance 2.5 rejects auto or returns resolution not priceable / aspect ratio not priceable Pass explicit resolution="480p" or "720p" , an explicit supported aspect ratio , and integer duration from 4–30. Seedance 2.5 uses token based pricing; call estimate cost(model, duration, resolution, aspect ratio) for a local estimate. MiniMax H3 reference request returns a parameter error Use image urls=[...] with the fal ai/minimax h3/reference to video model. generate video() translates it to upstream reference image urls ; do not hand send image urls to upstream. Job stuck IN PROGRESS 15 min Save request id , resume later with poll status.py User reports the fal.media link "shows nothing" / "blank page" Expected — fal serves with CSP: sandbox; default src 'none' . Deliver the local file at result["local path"] instead of the raw URL (see §1). 8. Infrastructure (reference) Caller → sc proxy → queue.fal.run (and api.fal.ai ) → fal model providers All requests must include Authorization: Key fake falai key 12345 (proxy injects the real FAL KEY ) Pre charge happens at submit. Poll/result calls are free. Allowed endpoints: video text to video / image to video / video to video / edit video for the registered models. Anything else returns 403 endpoint not allowed . Final mp4 lives at https:// .fal.media/... — public CDN, no auth needed for download. 9. Maintenance Adding a new model → register price in generate video.py::estimate cost and in transparent proxy/apis/falai.py:: VIDEO PRICING . Asset hosting via fal storage upload is intentionally not used in this skill: the production FAL KEY lacks Serverless permission. Keep using the preview based approach until that changes.