magic-hour-character-consistency
Keep a recurring character recognizable across Magic Hour images, storyboards, and video shots. Use for a mascot, fictional character, or authorized person appearing in multiple scenes; not a single portrait or face swap.
- License 许可
- MIT
- 最近更新
- 2026-11-14
- Stars
- ★ 1.6k
- Forks
- 491
01 前言 Front Matter
- name
- magic-hour-character-consistency
- license
- MIT
- metadata
- { "author": "magichourhq", "version": "1.4.0" }
- description
- Keep a recurring character recognizable across Magic Hour images, storyboards, and video shots. Use for a mascot, fictional character, or authorized person appearing in multiple scenes; not a single portrait or face swap.
02 资源文件 4
- 参考文档agents/openai.yaml298 B
- 许可证LICENSE1.1 KB
- 参考文档references/design.md4.5 KB
- 参考文档references/series.md2.4 KB
03 SKILL.md 正文
Magic Hour character consistency
Make the same character survive changes of scene, pose, and camera. A repeated name or seed does not establish visual identity.
Reuse the existing character and series context. Choose staging and technical settings yourself when delegated; do not make the user fill out a character template or repeatedly approve in-budget work. Respect requested review points and show the accepted scenes first. For a cast, new angle or recurring episode, read series continuity.
For an original recurring character without an accepted design, read character design.
Establish the anchor
Read the existing character sheet or approved image before asking for another. Separate fixed traits (face shape, hair silhouette, build, distinctive marks, rendering style) from scene variables (pose, expression, setting, lighting). Clothing is fixed only when the brief requires continuity. Record visible facts; do not infer hidden details or a real person's identity.
For a named saved character, call list_saved_items (or authenticated GET /v1/saved-items through the existing API) with type: character. Match the name and paginate with next_cursor as needed; disambiguate multiple matches. Inspect its primary IMAGE asset and any additional views the shot requires. Reuse their durable file_path values in compatible image fields, not the item ID or temporary preview URL. Keep an explicitly supplied anchor authoritative. Check the library before deciding no reference exists.
If there is no reference, create one useful view with generate_image. Prefer a clear face and readable silhouette at the scale the story needs. An elaborate turnaround is unnecessary for two similar shots; a back view or unusual angle may need its own reviewed anchor. Keep only individually accepted views—an inconsistent contact sheet is not a better reference.
Build shots from the original
For each scene, use edit_image with the approved identity image in assets.image_file_paths. If using additional pose/style images, describe each reference's role and check the chosen model's input limit. Do not merge a style donor's face or clothing into the character. Never claim reference ordering or weighting controls that the live schema does not expose.
Start with the scene change and a request to retain the reference character. State critical identity details when the brief needs them, but do not automatically copy the entire character sheet into every prompt. A detailed prompt does not rule out duplicates or omissions. In the Nori comparison, the shorter prompt preserved details better in one scene; added identity instructions helped another. Choose by the output, not prompt length.
Place this robot in a sunlit greenhouse, watering a small potted seedling. Keep the same character and watercolor-and-ink illustration style. Full body, three-quarter view. No text or extra characters.
Adapt that example to the actual reference and requested scene; do not add the robot to an unrelated request. Derive each new scene from an approved anchor, not from the last unreviewed scene. Compare the first new scene before spending on the rest.
For requested animation, review each shot's still first, then pass that exact image to create_video_from_image. Prompt for one coherent action and restrained camera motion. If a required face turns away or becomes small, inspect that interval particularly carefully. Match continuity at cuts: screen direction, eyeline, scale, wardrobe, palette and time of day. Do not promise identical pixels across generations or a seamless transition between independently generated shots.
Run and recover
Use the connected creation MCP at https://mcp.magichour.ai/; setup. Otherwise consult the live API schema: /v1/ai-image-generator, /v1/ai-image-editor, /v1/image-to-video. Choose supported parameters and count anchor creation, scene edits and animation against the existing budget. Never buy credits or expand a batch automatically.
Upload local files with get_upload_urls, PUT raw bytes to its upload_url without the API bearer token, then pass the returned file_path. Submit once, retain each ID and use wait_for_image_project or wait_for_video_project. Resume the same job after a polling timeout; do not repeat an ambiguous creation call. Preserve signed download URLs exactly and save durable outputs.
Accept and reuse
Compare all accepted scenes side by side with the anchor, not just each other. Check face/eye shape, proportions, distinctive marks, clothing continuity, hands, duplicate limbs and style. Count repeated details such as scarf ends and antennae; a recognizable character can still gain or lose them. Mark required details hidden by a prop or pose as unverified. Inspect the full videos for momentary identity changes. Reject only the failed scene and return to the anchor; do not redesign accepted scenes to match an error.
Deliver individual shots in order plus the accepted anchor and a short reusable identity note. Record changed traits and rejected outputs explicitly. Keep project IDs and local files so the next episode can reuse approved work without reconstructing it from chat. The Nori case publishes five authenticated MCP outputs and mixed prompt-comparison results; broader character coverage and animation quality remain unverified.
04 同仓库更多 Skills
worktable-review-with-annotations
Create, inspect, reply to, update, or resolve situated Worktable annotations. Use when feedback belongs to a Doc section, block, text range, or HTML Doc, or when the user asks to address existing review feedback and coordinate a requested artifact edit.
worktable-find-and-synthesize-context
Find, assess, and synthesize context from Worktable Docs and Records. Use when the user asks what a workspace contains, wants workspace-backed research, needs current facts reconciled across artifacts, or requests a cited synthesis without changing workspace state.
worktable-create-or-update-docs
Create, edit, rename, or maintain narrative Worktable Docs. Use for plans, decisions, briefs, notes, research syntheses, and other durable prose when the user requests a new artifact or a safe update to an existing one.
worktable-create-or-manage-records
Organize repeated structured information as durable Worktable Records. Use for research datasets, inventories, directories, pipelines, trackers, and other collections whose items need identity, fields, filtering, links, validation, or individual updates.
worktable-create-or-manage-interactive-html
Create or improve durable visual and interactive artifacts in Worktable. Use for dashboards, trackers, calculators, mini apps, slide presentations, and other requests that benefit from a polished HTML Doc.
worktable-collaborate-threads
Read, answer, and hand off messages through durable Worktable threads with users or other authenticated participants. Use when the user asks to check addressed messages, reply to a participant, continue a thread, request another agent's help, wait briefly for a response, or watch for comments and messages sent to the agent.