Velocity Skill · skill pública de Open Skill
Velocity Skill
USE THIS SKILL FOR PHOTOREALISTIC RESULTS WITH LOW COST MEME, ONE OF THE MOST BUDGET FRIENDLY MODEL FPR VIDEO GENERATION
Film
11 2 12
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/1502836/2026-09-09/hnrwx4fw.png
# Velocity Skill ## When to use - The user wants a photorealistic, cinematic AI video but wants to keep credit/cost down. - The user describes a "premium opening + economical middle + polished ending" strategy, or asks to split a video into bookends and a middle to save credits. - The user wants consistent characters/lighting/motion across a multi-segment video and is willing to lock a character reference first. - The user provides a script/idea and/or character/scene reference images, and optionally a voiceover or music track. - The user wants a video of any length (e.g. 10s, 30s, or longer), created in one go or extended from a shorter one. ## Core strategy (what this skill does) A video of **any length** is split into three segments with different cost profiles, then stitched into one seamless clip: - **Opening (premium bookend, 2s):** a striking, high-impact hook. - **Middle (economical bulk, total length − 4s):** the main action, generated on the default economical channel to save credits. - **Ending (premium bookend, 2s):** a polished payoff. The **creator decides the total length and the split — there is no fixed limit.** A 10s video uses 2s + 6s + 2s; a 30s video uses 2s + 26s + 2s; longer is fine too. The creator can also create a shorter video first and then extend it, or directly create a longer one in one pass. Consistency is guaranteed by **locking one character/look reference image first**, then feeding that same image into every segment so character, lighting, and motion feel continuous. The final result is one continuous video of the requested length. > **Important:** The platform maps stable *channels* (e.g. `fl2v`, `i2v`, `ref2v`) to the current best models automatically — you must **not** promise or hard-code physical model names like "Seedance" or "Grok". The cost saving comes from the *strategy* (short premium bookends + long economical middle), not from naming a specific model. If the user asks for a specific model by name, explain that the platform picks the best current model per channel and the strategy still delivers the budget benefit. For videos longer than 10s, the opening and ending use the premium bookend channel (`fl2v`) so the first and last seconds get the highest-quality treatment. ## Steps ### 1. Clarify input Ask for (or confirm) the following before generating anything: - **Total length and split:** ask the creator for the total duration (no limit) and how to split it. Default split is **2s opening + (total − 4s) middle + 2s ending**; the creator can adjust any of these. - **Script/idea:** what happens in the video (a one-liner is fine). - **Reference images (optional):** any character/scene images the user wants to use. - **Orientation:** default `portrait` (9:16, short-form); use `landscape` (16:9) if requested. - **Audio (optional):** whether the creator will provide a voiceover or music track, or wants one generated. - Break the idea into 3 beats: opening hook (2s), middle development (the rest), ending payoff (2s). ### 2. Lock the character / look (consistency anchor) Create a single clean, locked character reference image. This is the anchor for every segment. - **If the user provided reference image(s):** use `generate_image` with `model='i2i'` (or `i2i.hd` for higher quality) and `input_urls=[the reference image(s)]` to produce a clean, front-facing character sheet with consistent lighting and a neutral pose. Optionally run `recognize_image` on the refs first to describe the character accurately. - **If text-only:** use `generate_image` with `model='t2i'` to create the locked character from the description. - Set `name="Locked Character"`, `entity_type="characters"`. Keep this image and reuse it as input for all later steps. ### 3. Generate the opening (premium bookend, 2s) - From the locked character, generate two frames with `generate_image` `model='i2i'`: - an **opening first frame** (hero establishing shot, striking composition) — `name="Opening First Frame"`, `enti