Video Gen Usage

AI Video Generation, Transitions & Editing

• generate_video — Generate a video from a text description, optionally animating a source image (image_path). Saves an MP4 to the workspace and shows an inline player.
• generate_transition — Bridge two existing clips with an AI-generated transition. Returns the BRIDGE CLIP ONLY.
• edit_video — Edit an existing video with a text instruction (add/remove objects, restyle, relight, wardrobe, camera).

Model selection

• veo-3.1-fast (default) — exact 4/6/8s clips, native audio, 1080p/4k, 9:16 vertical, ~$0.10/s (4k $0.30/s). In our hands-on testing this is clearly the best quality per cost — use it for most generations.
• veo-3.1 — same controls at ~$0.40/s (4k $0.60/s), higher fidelity. Re-generate keepers on it when the deliverable demands top quality.
• omni-flash — 720p only, picks its own length (duration_seconds is a target, max ~10s), ~$0.10/s. Noticeably weaker visuals than Veo in our testing — use only when the flexible clip length is specifically wanted.
• Transitions — two complementary modes:
  • seedance (default, ~$0.18/s) watches both source clips and invents a creative transition — the impressive pick, but not pixel-continuous and occasionally cuts mid-bridge. Source clips are capped at 15s/50MB combined, so trim to the junction segments first (video-tools) when inputs are longer.
  • veo-3.1-fast ($0.15/s) / veo-3.1 ($0.40/s) animate between the first clip's last frame and the second clip's first frame — seamless by construction (both joins are guaranteed to match, no internal cuts). Use these when clean continuity matters more than spectacle, or as the fix when a seedance bridge won't stop cutting. With a FAL key configured these run via fal.ai and work everywhere; without one they fall back to Google direct, which region-gates first/last-frame ("use case not supported"; EEA keys affected).

Guidelines

• Describe videos like a shot brief: scene, subject, motion, camera work, lighting, mood, and what should be heard (all models generate audio). Vague prompts produce generic results.
• These are paid per-second calls — pick the shortest duration that serves the purpose, iterate on veo-3.1-fast or omni-flash, and only use 1080p/4k when the deliverable needs it.
• Generation takes 1–6 minutes per clip; this is normal — don't retry a call that is simply still running.
• generate_transition returns only the bridge clip. Assemble first clip + bridge + second clip with video-tools (declarative timeline, proper re-encode); never try to concatenate mismatched MP4s by hand. Set aspect_ratio: "9:16" for vertical footage.
• Judge the bridge before assembling — is it worth what it cost? If the generated transition reads as a plain dissolve between the two scenes (Veo first/last-frame tends to when the scenes are very different), don't pay to regenerate: assemble with a free video-tools fade/dissolve instead. AI bridges earn their cost when generated motion adds something a crossfade can't — a camera move, a morph, an invented traveling shot. Veo shines on same-scene/continuation joins; seedance on creative scene changes.
• Verify the joins before delivering an assembled transition. Compare the bridge's first/last frames against the neighbouring clips' junction frames (video-tools render_frames). Veo bridges are frame-anchored, so prefer HARD CUTS at the joins — but trim the bridge's duplicated endpoint frame first (its first frame ≈ clip A's last frame; showing both holds the image for 2 frames and reads as a stall). If a residual color/exposure or framing mismatch survives, fix the mismatch — grade the bridge to match its neighbours (video-tools) — rather than hiding it; a short crossfade at the joins is the last resort (and the norm for seedance, which is not pixel-continuous). A style prompt asking to "end exactly on the framing of the second clip's opening" also helps.
• Verify the bridge has no internal hard cut. Seedance is natively multi-shot and sometimes "transitions" by cutting mid-clip — render a few frames across the bridge and check the motion is continuous. If it cut, regenerate with a style prompt that hammers "one unbroken shot, no cuts" (the built-in template already demands this, but the model occasionally ignores it) — or switch to a veo-3.1-fast transition, which is frame-anchored and cannot cut.
• edit_video accepts 2–30 second inputs up to 1080p (~$0.28/s of output). Pass duration_seconds (the input clip's length, from probe_media) so the edit is priced correctly in the platform's cost tracking. For longer videos, cut out the segment to edit with video-tools, edit it, then splice it back. Aleph returns SILENT video — re-mux the original audio with video-tools when sound matters.
• Generated MP4s land in the workspace under generated-assets/ (a bare filename via save_path also lands there). Use a relative path with a subfolder (e.g. "projects/launch/hero.mp4") to organize into your own folder.
• The saved MP4 is web-safe: the inline player is pushed automatically; feed the file to video-tools for assembly, captions, or grading.