Faithful Video Generator (i2v & ref2v) · عوامی Open Skill
Faithful Video Generator (i2v & ref2v)
Faithful Video Generator (i2v & ref2v) Digen پر ایک عوامی Open Skill ہے۔ کھولیں اور دیکھیں یہ ورک فلو کیا تخلیق کرتا ہے۔
Film
1823 82 255
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4130853/2026-09-16/2tcmszx6.jpg
# Faithful Video Generator (i2v & ref2v) ## When to use - The user uploads a single image and wants it animated / turned into a video: use `i2v`. The uploaded image is the **start frame** and the **only visual basis** — never invent or add any new elements, characters, backgrounds, props, or scenery that are not already in the image. - The user provides one or more reference images and describes a video (character / subject consistency): use `ref2v`. Use exactly the reference images the user provided, in the role and scene they describe. - Only handle video generation. Do not generate images, voiceover, or music unless the user explicitly asks. - Generate a **silent 10-second clip by default**. Do not add talking / dialogue / voiceover, music, or any other audio unless the user explicitly asks for it. ## Core principle (mandatory) - **Strictly faithful to the user.** Pass the user's own prompt to `generate_video` essentially verbatim. Do NOT embellish, expand, rewrite, or inject your own creative ideas. Only fix obvious typos/grammar if needed, and keep the meaning identical. - **i2v = image is the base and start frame.** Never add new visual content; the motion/action must stay within what the image already shows. If the user's prompt would require a new element not in the image, follow only what is possible from the existing content and note the limitation to the user. - **ref2v = reference images as-is.** Pass the user's reference image(s) as `input_urls` and let the user's prompt describe the scene/motion. Do not swap, restyle, or reinterpret the subject. - Never generate content that isn't requested, and never invent a subject that isn't in the user's image/prompt. - Do not add talking, voiceover, or music. Only include dialogue or background music if the user explicitly asks for it in their prompt. - **Anti-artifact quality guidance (always append, for both i2v and ref2v).** Append short stabilizing descriptors to the prompt so the output stays natural and faithful: the clip's motion must begin gradually from the exact first frame (no instant / sudden jump of the subject at the very start); preserve the subject's hair exactly as in the image — natural hair strands and edges, never blurred, melted, or smoothed into a plastic look; keep skin and all textures natural, with no over-smoothing, waxy / airbrushed "AI" finish, no beauty filters, and no restyling. This guidance only stabilizes rendering quality; it must never add, remove, or change the content of the image or the user's request. ## Steps 1. Read the user's message: which image(s) they uploaded and exactly what they want the video to do. Do not assume or infer additional creative details beyond their words. 2. Choose the channel: - Single uploaded image to be animated → `model='i2v'`, `input_urls=[<that image asset_id>]`. - One or more reference images for identity consistency → `model='ref2v'`, `input_urls=[<the reference asset_id(s)>]` (at least 1 reference image required). - Audio-driven reference video (only if the user asks): keep `model='ref2v'` and pass `audio_url` (optionally `audio_start_time`) — do not put the audio into `input_urls`. 3. Call `generate_video` with: - `model` = the channel chosen in step 2 - `input_urls` = the user's image asset_id(s) (copy the short `asset_id` from the upload or a prior result verbatim; never guess or reconstruct a URL) - `prompt` = the user's own prompt, kept faithful (see Core principle), plus the always-appended anti-artifact guidance (gradual motion starting from the exact first frame; natural hair strands and edges with no smoothing/blur; no AI filters or overprocessed finish) - `orientation` = the aspect the user asked for; if not specified, default to `portrait` (9:16) - `resolution` = `720P` default - `duration` = `10` seconds (the base clip length; increase only if the user asks for a longer video) - `name` = a short descriptive title of the clip - `entity_type` = `final_video