vincevangen
Sign in
Docs menu

Video

Writing video prompts and choosing a tier.

This is the same playbook the agent skill carries. Names like fast and default are aliases; see Models for what they run on the hosted service; model names in the research notes below may include ones only the local tool offers. Prices are in credits.

Formulas [vendor]

  • Wan: Entity + Scene + Motion; advanced: Entity(desc) + Scene(desc) + Motion(desc) + Aesthetic control + Stylization. Image-to-video: Motion + Camera only, the image fixes the rest. (Alibaba Model Studio guide.)
  • Veo 3.1: [Cinematography] + [Subject] + [Action] + [Context] + [Style & Ambiance]; "prompt like you're directing a shot". Audio: A woman says, "…", SFX: thunder, Ambient noise: …. Timestamps [00:00-00:02].
  • Kling: Subject + Subject Movement + Scene + (Camera + Lighting + Atmosphere). Movement "suitable for a 5-second video". Simple words. Weak at complex physical movement and counts. 2,500-char limit on prompt and negative.
  • FLUX 3 video (BFL skill): opening frame → visible cause → physical response → payoff → continuity contract ("no cuts, transitions, frozen or repeated frames").

Rules

  1. Action budget: one subject action and one camera move per clip [vendor: BFL]. No cuts or scene changes inside a clip.
  2. Camera always explicit. Wan defaults to a mild push-in when no camera is given [community]; write "static locked-off shot" if you want none.
  3. Order: subject action (as it unfolds) → camera → environment motion → pace [community]. Spell out the physics; if there is no motion add ambient motion (wind, steam, drifting clouds) [vendor: Wan extender].
  4. Style must be visible: "tense" becomes rigid posture, hard side light and a slow push-in [vendor: BFL]. 2D styles get no film-camera terms [vendor: Wan].
  5. Lint contradictions: a locked-off camera cannot track; an extreme close-up cannot establish [vendor: BFL].
  6. Length: 60-200 words (Wan extender) [vendor]; simple sentences.
  7. Do not rely on legible text, logos, many-item counts, lip-sync to specific words, or choreography; add typography and logos in post.
  8. Draft the first frame as an image, approve it, then animate with image-to-video. Cheaper and more controllable [inference]; Veo's first-and-last-frame workflow does the same [vendor: Google]. Locally video:fast (p-video) animates an image: frame 0 is the image, a transparent sprite comes back on an opaque background [measured 2026-10-12]. Hosted, video:fast is text-only: use video:i2v.
  9. After two structurally similar misses, change the concept or the production method, not the adjectives [vendor: BFL].

Wan default negative prompt (pass verbatim on Wan models) [vendor: Wan2.2 shared_config.py]

色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走

Meaning: garish colours, overexposed, static, blurry detail, subtitles, grey cast, worst/low quality, JPEG artifacts, ugly, extra fingers, bad hands/faces, deformed limbs, fused fingers, motionless frame, cluttered background, three legs, crowded background, walking backwards.

Vocabulary [vendor: Wan]

Shot sizes: extreme close-up, close-up, medium, full, extreme full. Angles: eye-level, over-the-shoulder, high, low, top-down, aerial. Moves: push-in, pull-out, tracking, orbit, fixed. Light: soft, hard, side, rim, backlight. Time: dawn, dusk, golden hour, night. Tone: warm/cool, high/low saturation.

DeepInfra video specifics

Video models return video_url as a data:video/mp4;base64 URI (tens of MB) or, for Veo, a videos list [vendor: DeepInfra docs]. Runs can take minutes (one documented Wan2.2 run: 267 s). vincevangen submits with a webhook or polls; MCP callers get a job id and use vvgen_get_job.