fal-seedance-video

name

fal-seedance-video

description

Implement and operate fal.ai video generation for AI UGC clips, with Kling O3 Standard as the default cost-effective model and Seedance 2.0 as the higher-cost quality baseline. Use when generating or coding TikTok/Reels/Shorts videos with fal video models, choosing endpoints, setting aspect ratio/resolution/duration, uploading image references, submitting queue jobs, polling results, handling webhooks, or producing fal request payloads for AI UGC scene clips.

fal Video Generation

Use this skill when Codex needs to generate or implement video clips with fal.ai video models. Default to Kling O3 Standard for cost-effective UGC tests, and use Seedance 2.0 when you need the higher-cost quality baseline.

Read references/fal-seedance.md when you need Seedance endpoint IDs, parameters, payload examples, or cost/resolution guidance. Read references/fal-video-providers.md when comparing Seedance against cheaper fal providers.

Defaults for UGC ads

• Default provider: Kling O3 Standard image-to-video.
• Quality baseline provider: Seedance 2.0 image-to-video.
• Aspect ratio: use vertical 9:16 source images for Kling; set aspect_ratio: "9:16" for Seedance.
• Resolution: set 720p for Seedance to control cost; Kling O3 Standard does not expose the same resolution field.
• Duration: 5 or 6 seconds for most UGC chunks; 10, 11, or 12 seconds for longer mechanism/payoff chunks when supported.
• Safety checker: leave enabled.
• Camera: keep subtle and realistic; avoid aggressive cinematic movement for talking-head UGC.

Provider candidates

• Default cost-effective model: fal-ai/kling-video/o3/standard/image-to-video
• Higher-cost quality baseline: bytedance/seedance-2.0/image-to-video
• Low-cost visual-only exploration: fal-ai/minimax/video-01/image-to-video

Workflow

1. Confirm an API key is available before running generation. For AutoVisuals UGC work, use UGC_FAL_API_KEY so creator/ad generation does not share app provider keys. For general projects, FAL_KEY or FAL_API_KEY are also supported.
2. Choose endpoint:
   • fal-ai/kling-video/o3/standard/image-to-video for default lower-cost image-to-video UGC tests with generated audio.
   • bytedance/seedance-2.0/image-to-video for higher-cost UGC clips from persona/reference images with generated synchronized audio.
   • fal-ai/minimax/video-01/image-to-video for cheap visual motion tests where native dialogue is not the focus.
   • fal-ai/bytedance/seedance/v1/pro/fast/image-to-video for older visual-only or legacy workflows.
   • fal-ai/bytedance/seedance/v1/lite/image-to-video for cheaper fast prototypes.
   • Text-to-video only when no reference image is required.
3. Upload local images with fal.storage.upload or provide public image URLs.
4. Submit each chunk as a separate request.
5. Store request IDs, prompt, input image, duration, resolution, aspect ratio, and output URL.
6. Review every output before stitching.
7. Regenerate only failed chunks; preserve working clip outputs.

Local helper

Use scripts/generate_fal_video_clip.mjs for one-off local clip generation when the current project has @fal-ai/client available or can install it.

Example:

UGC_FAL_API_KEY=... node /Users/jjeaton/.codex/skills/fal-seedance-video/scripts/generate_fal_video_clip.mjs \
  --provider kling-o3-standard \
  --prompt-file /path/to/chunk-01-prompt.txt \
  --image /path/to/marc-reference.png \
  --duration 6 \
  --generate-audio true \
  --out-json /path/to/chunk-01-kling-result.json

If a project should own this permanently, copy the helper into the project and add @fal-ai/client as a project dependency rather than relying on the global skill path.

Failure handling

• 422 client errors may be charged. Validate payloads before sending.
• If product/reference appears in a no-product scene, remove the asset mention entirely and regenerate.
• If voice pacing is too fast, increase duration or split the chunk.
• If voice pacing is too slow or identity drifts, shorten duration and regenerate.
• If hands/product deform, simplify the hand action and keep product stationary or out of frame.
• If face identity drifts, use image-to-video with a stronger reference frame and less camera movement.