HaoyangZhang3/ppt-xqmsav-skill

AI One-Shot Morphing Transition Video

Overview

Create an "one-shot" style AI morphing transition video: multiple landmark scenes seamlessly morph into each other in a continuous sequence (element from scene A transforms into an element of scene B), ending back at the starting point to form a loop, all unified under a high-contrast color scheme (e.g., red/white/black poster style). The final product is commonly used for PPT openings, city promotional videos, or viral account content. See the complete Hangzhou case study's four stylized reference images in assets/case-hangzhou/.

Workflow (Five-Step Method)

Step 1: Design the Scene Chain

• Select 4~6 landmarks/scenes from the same city, arranged in a circular structure: the start is also the end (e.g., Leifeng Pagoda → Lingyin Temple → Hangzhou Arch/Lotus Bowl → Tensioned Membrane Bridge → back to Leifeng Pagoda).
• Sorting principle: there must be a shape/color/structure similarity point between every two adjacent scenes (see Step 3). Draw the chain diagram before starting.
• Determine the unified visual style and color palette (case study uses red/white/black flat poster style; alternatives include green-blue landscape, cyber neon, etc., but the entire video must use one consistent set).

Step 2: Unify Source Images into One Style

• Collect real photos of each landmark. Use Jimo's "Image Generation - Reference Image to Image" (or similar tools) to convert all photos into the same style: uniform color scheme, consistent brushstroke, silhouetted, flat.
• This is the critical success factor: if styles are not unified, the morphing between segments will look unnatural and broken.
• Output 4~6 stylized images, numbered and saved (Image 1, Image 2...). These will serve both as reference images for video generation and as the "first frame/last frame" for each segment.

Step 3: Design Match Cut Points

For each pair of adjacent scenes, write a "morphing correspondence relationship". Priority order: shape similarity > same color > repeated structure. Case study comparison table:

Previous Scene Element	Next Scene Element	Morphing Verb
Red sky + flying birds	Red buildings	Land as
Lotus Bowl white petals	Tensioned membrane white framework	Roll, rearrange
City black road	Black water surface	Extend into
Red road with white markings	White boardwalk	Converge into
Red membrane roof	Red water surface	Pour, spread out
Boardwalk white planks	Arch bridge white arch opening	Close, curve

See references/prompt-patterns.md for the complete verb library and sentence templates.

Step 4: Write Prompts and Generate Segments

Use video models that support "multi-image reference / first-last frame" like Jimo Video or Keling:

• Method A (First-Last Frame Relay, recommended for long takes): For each clip, set "first frame = Image A, last frame = Image B". The prompt only describes the morphing action from A to B. Break a long chain into multiple segments, relay by using the last frame screenshot of the previous segment as the first frame of the next segment, ensuring smooth transition.
• Method B (Long Prompt with Multi-Image Reference): Within a single prompt, reference images like @Image1, @Image4 as start and end scenes, following a chain description of "scene → action → landing point" (see references for the Hangzhou case original text).
• Fixed prompt structure: [Subject] + [Directional Action] + [How the Shape Morphs] + [Corresponding Element in Next Scene] + [Camera Movement].
• Camera movement verb library: fly in, dive, pass through, fall, lift, flip, vertical top-down shot, pull out along water surface, climb over, tilt up.
• Each segment: 5~10 seconds; do not include more than 2 morphing actions per segment otherwise it will likely break.

Step 5: Post-Production Assembly

• Use a video editor (e.g., CapCut) to splice segments in sequence (segments generated via the first-last frame relay method will be nearly seamless).
• Add landmark subtitle bars (e.g., "LINGYIN TEMPLE"), and sync with drum-heavy background music.
• Use scripts/extract_frames.py on the final video to extract frames at intervals and check for broken transitions; regenerate the problematic segment if necessary.
• Export in vertical 9:16 (or 16:9 for PPT openings), and insert into the first slide of your presentation for autoplay.

User Interaction Points

• When user only provides a city name: First list 4~6 landmarks for that city, ask questions to confirm the scene order and color style, then proceed to Step 2.
• When user provides real photos: First confirm the stylization direction (present 2~3 color scheme options).
• When user requests "the same style as the Hangzhou case": Directly show the four images from assets/case-hangzhou/ and the original prompts from references; replace the city name to adapt.

scripts/

• extract_frames.py — Uses OpenCV to extract frames from a video at regular intervals and assembles them into a nine-grid inspection image with timestamps, used for checking transition seamlessness. Usage: python extract_frames.py <video.mp4> [output_directory] [number_of_frames].
• palette_unify.py — Unifies a batch of generated images into the red/white/black poster palette, eliminating sky/ground color differences between images to prevent color jumps during transitions. Usage: python palette_unify.py <in.png> <out.png>.
• palette_recolor.py — Globally shifts the hue of poster images according to a preset scheme (bluegold cyber blue-gold / darkred dark wine red). Note: For dark scene styles, prefer re-generating the images with AI; local brightness remapping cannot distinguish between "white sky" and "white buildings" which can easily overexpose into white blocks. Details in references/color-themes.md.
• assemble_video.py — Final video assembly: crops out model watermarks, adds cross-fade between segments to remove seams, overlays a main Chinese title and landmark subtitles. Usage: python assemble_video.py <out.mp4> <seg1.mp4> [seg2.mp4 ...] [--w 1920] [--h 1080] [--accent gold|white|red]. The --accent flag controls subtitle color (gold=champagne gold, white=white, red=wine red).

Color Themes

Three established themes (Flat Poster Red/White/Black / Dark Cyber Blue-Gold / Dark Wine Red) with full style words, color proportions, and style-changing action lists are available in references/color-themes.md. Changing the style = re-generating 4 keyframes + replacing color words in 4 prompts; this is low-cost (about 30-40 credits for 4 images) and yields far better results than local color shifting.

Generative Tool Parameter Quirks (Tested)

• The video model only accepts 768P / 2K resolution; requesting 1080p will cause a direct error.
• The video model forcefully generates an audio track; passing enable_audio: false will cause an error. Simply omit the parameter (this field is only supported by some models).
• The generated video's "AI generated" watermark is in the top-left corner (it appears in the bottom-right for static frames). When assembling, you need to crop the top and bottom edges accordingly.
• Occasionally, the same prompt will return a requires prompt error; simply retry.

Resources

references/

• prompt-patterns.md — Prompt sentence templates, morphing verb library, camera movement verb library, and the complete original prompts for the Hangzhou case. Must-read before writing prompts.

scripts/

• See above.

assets/

• case-hangzhou/ — Four stylized reference images from the Hangzhou case (Tensioned Membrane Bridge, Leifeng Pagoda, Hangzhou Arch Lotus Bowl, Red Sky Leifeng Pagoda), serving as style and composition examples.
• case-ncepu/ — Four stylized reference images from the NCEPU Beijing campus case (Arch Bridge Loop, Red Sky Main Building, Campus Panorama, Red Membrane Framework), which can serve as a composition template for "campus/city landmark" type requests.