MUSIC VIDEO CREATOR LOW CREDITS · skill Open Skill công khai
MUSIC VIDEO CREATOR LOW CREDITS
MUSIC VIDEO CREATOR LOW CREDITS là một Open Skill công khai trên Digen. Mở ra để xem workflow này tạo gì.
Music
21 4 19
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/2175014/2026-09-09/ecphfdeq.jpg
# Song to Music Video Build a 16:9 music video from an uploaded track. Default spec is 24 shots × 10 seconds ≈ 240 seconds. Present styles first and wait for a choice unless the user already named one. ## Capability reality Do not call tools that do not exist in this environment (transcribe_audio, platform generate_video, mix_clips, ai2v, i2v, t2i.hd). Use this stack: - Lyrics or structure — user-pasted lyrics, or a short structural guess from duration and mood if instrumental. - Stills — generate_image (file on disk for ffmpeg) or render_generated_image (chat preview). Prefer generate_image with landscape orientation so files land under artifacts/imagine_images/. - Motion — ffmpeg Ken Burns / slideshow (load the ffmpeg skill). Real-Motion 3.5 is not connected. - Mix — scripts/assemble_music_video.sh muxes stills plus the uploaded audio. State this fallback once if the user asked for Real-Motion 3.5 or ai2v. ## Style choice (required before generating) Present these options as a numbered list and ask the user to pick a number, a name, or a custom description. Do not generate stills until a style is chosen, unless they already specified one in the same message. 1. Cinematic Film — anamorphic widescreen, film grain, natural light, shallow depth of field 2. Anime — hand-drawn look, vibrant colors, dramatic lighting 3. Cyberpunk Neon — night city, neon signs, rain reflections, high contrast 4. Dreamy Ethereal — soft focus, pastel palette, floating particles, surreal 5. Retro 80s — VHS grain, synthwave palette, chrome, sunsets 6. Dark Fantasy — epic landscapes, fog, dramatic skies, muted tones 7. Minimalist Abstract — geometric shapes, clean gradients, slow elegant motion 8. Documentary Nature — wildlife, landscapes, natural light, realistic 9. Custom — the user describes palette, era, medium, camera, and mood in their own words If they pick 9 or write a freeform style, lock that text as the style block and reuse it verbatim in every image prompt. Read references/styles.md for prompt fragments. Default only if they explicitly say they do not care — then use Cinematic Film. ## Inputs Required: - Uploaded audio (mp3, wav, m4a, aac, flac, ogg). Stop and ask if missing. Optional: - Style (preset 1–8 or custom). - Pasted lyrics (strongly preferred for a lyric-faithful board). - Reference images labeled as character / scene / prop. Keep appearance consistent. Cap useful refs at a small set; prioritize character, then scene, then prop. - Length override. Default 240s. If they want 60s, use 6 shots × 10s or 8–12 shots as in lyrics-music-video. If the track is longer than the target, cover the first N seconds. If shorter, end when the song ends and reduce shot count so each shot stays about 10s. ## Workflow ### 1. Confirm spec Probe the audio with ffprobe. Confirm style, duration, shot count, and 16:9. Pure visual storytelling (no singer lip-sync) unless the user asks for a performer. ### 2. Structure the song If lyrics exist, map them onto an arc. If not, use duration only — intro, build, hook, verse, second hook, bridge, climax, outro — scaled to shot count. ### 3. Write the shot board For shot i at 10s each, record scene, camera move, mood, and still prompt. One coherent world. No on-screen text or watermarks. For cost, if the user asked for the cheapest path, offer 8–12 shots before generating 24. ### 4. Generate keyframes Call generate_image once per shot, orientation landscape. Every prompt must include the locked style fragment, consistent character and grade, wide 16:9 frame, no text. Copy outputs into /home/workdir/artifacts/amv/shot_01.png and so on. If a shot fails, retry once with a slight rephrase, then continue. ### 5. Assemble Run scripts/assemble_music_video.sh with the stills directory, audio path, output mp4, and duration. Probe the file. Deliver with render_file. Preview a few stills in the reply. ### 6. Reply Include style used, shot count, duration, a short shot list, honest resolution (