Giant Avatar Consecutive Frame to Video
Break a readable power display into sequential states, then hand off the motion between adjacent states to video generation. Deliver a plan, keyframes, or the final video as per the user's current request; do not automatically expand the scope.
Picking up from Current Progress
- Approved style reference: Use directly as the master template; do not re-select styles.
- Approved consecutive frames: Check for contradictions between frames before generating video; do not re-generate images.
- Only plan, skill, prompts, or keyframes needed: Complete the corresponding part; do not automatically submit a paid video task.
- When the user says "continue after selection," stop at the selection point; if already selected or authorized, do not reconfirm.
- If a non-critical preference is missing and autonomous decision is allowed, adopt a reasonable default and proceed. Rainy night, eight frames, 25 seconds, a sword with music are not fixed requirements.
Can accept event cards from novel-distill, or directly use the user's ideas; no additional skill installation required. Distinguish between original text facts (F), audiovisual processing (P), and new adaptations (A). Visual references like domains, avatars, and weapons from reference videos must not be backfilled as novel facts. When no original text exists, refer to it as a visual concept interpretation.
Tool Selection: Platform Agnostic
This skill has a fixed production method but no default final platform. Image generation, image-to-video, and editing/mixing can be done with different tools or within the same platform.
- If the user specifies a tool or has an existing project, prioritize using it, verifying its actual capabilities. Do not automatically choose Pexo just because it was used in a historical case.
- If unspecified, select based on currently available tools, reference image constraint capabilities, duration, downloadable output, and budget, providing a brief rationale. If a viable plan already exists and the user permits autonomous decisions, no additional confirmation step is needed for platform selection.
- If no usable generative tools are available, deliver independent keyframes, a general segment brief, and prompts, explaining the missing capabilities. Do not describe prepared materials as already generated video.
- First verify the current API or interface, then map the general brief into specific parameters. Platform names do not equal model capabilities; do not assume every platform supports multi-image reference, start/end frames, or native sound effects.
1. Locking the Co-Frame Relationship and Continuity
Use a brief continuity table to record existing information:
| Item | Content to Lock |
|---|---|
| Characters | Face, age, hairstyle, robe silhouette, material, identifying features |
| Avatar | Default is the user's magnified projection, same face and attire; if the user specifies a deity/alien form, record the correspondence |
| Co-frame space | User's body in the foreground, avatar behind, taller than buildings; both the small figure's silhouette and the giant's face readable simultaneously |
| Body and frame coordinates | Which anatomical hand holds what, where the start and end points are in the frame |
| Scene | Architectural landmarks, light sources, weather, camera position and action axis |
| Material evolution | E.g., silvery phantom → skin tone cloth → edge dissipation; do not impose if the user does not need it |
| Attack execution | Start → charge → release → recovery, including number of times, direction, affected objects, and environmental reactions |
Map the same action, not a horizontal mirror. The user's anatomical right arm corresponds to the avatar's right arm; do not write ambiguous "mirror pose" causing left-right reversal. The avatar must not become a second opponent in the distance; a close-up of the face alone cannot replace the required co-frame key image.
Default to a single camera position or slight push-in based on the master template to reduce continuity issues; when the user wants multiple camera positions, clearly define the cuts and action axis, do not turn a single position into a style prohibition.
2. Design State Changes First, Then Generate Independent Frames
Read references/keyframes.md. For each frame, record the user's pose, avatar pose/transparency, weapon state, environmental reaction, differences from the previous frame, and actions not yet taken. Differences should be visually present in the image.
A common structure is "charge → domain first opens → first appearance → expansion → consolidation → gather weapon and raise hand → simultaneous release → aftereffects," which can be merged or split. When the gap between "gather weapon" and "raise hand" is too large, add transition states; do not cram more requirements into the same prompt.
Use currently available image generation/editing tools; when the user specifies a platform or model, confirm the actual callable capabilities. If the API does not return a version, record the channel and "version undisclosed"; do not claim specific model names like GPT-Image 2.
Generate one independent frame at a time, maintaining the aspect ratio. Usually input "approved master + last approved frame": the master maintains identity, the previous frame maintains state. Correct errors in the previous frame first; do not propagate errors, and do not input all images at once. Early frames must suppress climax elements from the master: no giant face if not yet manifested, no complete sword light if not yet released.
Inspect each frame after generation before continuing. If only the avatar's arm is wrong, make a targeted correction, preserving the correct user body, face, clothes, and scene. Reasonable differences in local textures are allowed; major contradictions in identity, hand side, and action phase must be resolved first.
3. Assigning Inter-Frame Motion to Video Segments
Read references/video-and-qa.md. Segment based on duration capability and action complexity, with adjacent segments sharing a passed boundary state. Shared images do not guarantee natural continuity after generation.
For each segment, write the initial state, final state, process reference, main action, camera movement, correspondence between parties, environmental forces, sound beats, and ending pose. Images fix appearance and state; video prompts describe the motion between states; do not create a static image slideshow.
If there are dedicated start/end frame fields, use them according to the API; a plain image_list can only be called an ordered reference. When there is no end frame constraint, check the actual segment end; if necessary, use the passed segment end as the next segment's start frame. Only submit independent segments simultaneously when boundaries are clear and the tool supports the corresponding constraints.
Only read the optional adaptation guide references/pexo.md when Pexo is actually used this time. It is not the default entry or required dependency; for other tools, follow the general brief and current capabilities.
4. Inspection, Correction, and Composition
First, look at the actual generated material, then decide on regeneration. Prioritize fixing co-source relationships, hand side and direction, repeated actions, action regression, and major identity drift. Judge material deviations against the original target; a platform's auto-check text alone cannot trigger a paid retry.
Co-frame does not mean synchronized. If the user has sheathed their weapon while the avatar still holds it up, that is a critical error. Check the period before and after each release and every junction point; use reasonable cropping, matching cut points, and sound effect bridges to handle minor differences; do not use reverse playback or long dissolves to mask erroneous poses.
Record original and corrected versions; do not overwrite the only successful piece of material. Clearly define defects in each round; reserve retry and post-processing budget. If the user sets an upper limit, do not exceed it without permission; also report deviations from the assistant's own cost targets. Multiple corrections should be itemized, not lumped into a single "round" that incorrectly implies free retries.
Arrange sound according to charge, release, and aftereffects; preserve effective ambient and action sound effects. If the soundtrack is too long, cut and fade it out first; only consider regeneration if the content is unsuitable. "Replacing the audio track" cannot be claimed as "overlaying and preserving." If the platform cannot mix audio, external post-processing can be done under existing authorization, but clearly explain the division of labor.
Local technical inspection example:
python3 scripts/audit_video.py /absolute/path/final.mp4 --cuts 6.04,12.08,17.12 --out /absolute/path/qa
Cut points must be actual timestamps. Requires Python 3, ffmpeg, ffprobe. The script outputs media parameters, full timeline overview, extracted frames at cut points, black frame detection, and volume checks; it does not judge character synchronization, subjective audio quality, or presence of dialogue; those must be checked manually and via appropriate audio inspection.
Deliverable
Deliver independent keyframes and/or a downloadable final video according to the agreed scope, preserving the state table, prompts, and production records. Record the actual tools/confirmed model, timeline, retries, costs, and residual deviations. For the final video, also record source material and output specifications; do not claim upscaling as native high resolution, and do not mistake platform "completion" for acceptance.
Ultimately, directly present the work and any significant limitations. When real examples are needed, read references/worked-example.md; historical prices and results do not constitute a promise of success rate, budget, or follower growth.