Communitygithub.com

Drama Keyframe Forge

Before any video render, locks one style bible and a 16:9 keyframe per shot so faces, rooms, and props stop drifting.

What is Drama Keyframe Forge?

Step two of the bench. It takes the shot plan and the character sheets and builds a single style bible, covering genre, palette, light, and materials, plus a high-resolution first frame for every shot with the scene, the people, and the props in one image. Those frames are the cheap check. Video sampling starts only after the frames pass, so text-to-video is not guessing the face.

Works with~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/abserari/drama-production-skills/tree/HEAD/skills/drama-keyframe-forge

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

Drama Keyframe Forge (Storyboard Keyframe and Asset Consistency Process)

Process Positioning: Short drama asset and pre-keyframe process. Core Concept: Upgrade the inefficient workflow of "blindly running T2V for 8 minutes, hoping for the best, and relying on post-production quality checks and re-draws" to a systematic approach: "Before video sampling, use an image generation model to create storyboard keyframes in 5 seconds, cheaply verify and lock them in, then feed them to I2V for execution."


1. Operational Contract, Inputs, and Outputs

  • Inputs:
    1. plan.json from the 1st process and the storyboard table;
    2. Character Reference Sheet (2+3 head-body decoupled views);
    3. Scene settings and key prop descriptions.
  • Outputs:
    1. style_bible.json: The globally unique, immutable art style master card;
    2. 16:9 high-definition first-frame placeholder images for each storyboard: segments/segXX_keyframe.png;
    3. First-frame gate quality inspection sign-off sheet: keyframe_qc.json.

2. Hardcore SOP: Global Style Bible and Pre-Production Style Gate

Rule: The entire short drama must have a single, legitimate style_bible.json. Individual storyboard prompts are strictly forbidden from freely playing with or privately tampering with core style words!

2.1 Five Core Attributes of the Style Bible

Before starting work, the following 5 fields must be solidified:

  1. Visual Genre: e.g., Cinematic retro-future dieselpunk graphic novel style, mature neo-noir tone;
  2. Color Palette: e.g., desaturated amber and deep teal contrast, rusty bronze highlights, rich shadow details;
  3. Lighting Base: e.g., heavy chiaroscuro lighting, harsh directional rim light, high contrast dramatic shadows;
  4. Environment Base: e.g., weathered rust and scuffed brass, damp asphalt reflections, atmospheric smog;
  5. Quality & Lens: e.g., 35mm film grain, Panavision anamorphic lens flare, shallow depth of field, 8k resolution.

2.2 Pre-Production Style Verification Checklist

Before starting work, generate 4 style verification samples at different shot distances using the image generation model, and obtain manual sign-off:

  • Sample 1 (Long Shot - LS): Verify if the lighting, color tones, and material atmosphere match expectations;
  • Sample 2 (Medium Long Shot - MLS of the protagonist): Verify the protagonist's costume, posture, and signature accessories;
  • Sample 3 (Close-Up - CU of the protagonist): Verify eyes, face features, hairstyle, and skin texture;
  • Sample 4 (Extreme Close-Up - ECU of key prop): Verify the prop's hard surface details and mechanical structure.

3. Hardcore SOP: Storyboard-Level Combined Keyframe Image Generation (I2V Pre-Generation)

Counterexample (Strictly Prohibited): Directly feeding plain text to H3 for T2V without a storyboard first-frame keyframe image! Symptom: Due to random noise in the latent space during the early diffusion stages in 3D, the same generated character may have a changed hairstyle, facial bone drift (face swapping), or different clothing styles across shots.

3.1 Combined Keyframe Image Generation Steps

For each storyboard segment (seg01 to seg08) in plan.json:

  1. Assemble the storyboard image generation prompt:
    • Inject the protagonist's specific appearance and clothing;
    • Inject the camera position specified for the current segment (e.g., M2_RACK_FOCUS for a prop focus pull);
    • Inject the specific micro-scene and props for the current segment (e.g., sharp serrated hunting dagger vibrating in tabletop);
    • Inject the style and lighting terms from style_bible.json.
  2. Use a fast image generation model (Flux / SDXL) to batch render 16:9 static images (5~10 seconds per image);
  3. Triple Gate Quality Check (Keyframe Gate Check):
    • Face Consistency: Key features (eyepatch, hairstyle, scar) are clearly aligned;
    • Camera Obedience: Shot distance and angle strictly follow the storyboard requirements, no extra body parts;
    • Prop Hard Surface: Edges of weapons, dials, and suitcases are sharp, with no melting;
  4. After passing inspection, rename and save the files as segments/{seg_id}_keyframe.png.

4. Dedicated Tool Scripts

  • Manage Style Verification and Storyboard First-Frame Generation: python scripts/keyframe_gate.py --scan
  • Sync First-Frames to ComfyUI Storage: python scripts/keyframe_gate.py --project <project_name> --sync

Individual skills in this repo

This repo contains 5 individual skills — each has its own dedicated page.

comfy-h3-render-core

短剧工业化管线·第 3 工序:H3 物理视频渲染与防翻车调度。负责与本地 ComfyUI 服务交互,调度 MiniMax H3 进行 Turbo 8-step 视频生成。硬核沉淀:Windows WDDM 驱动锁防御、潜空间 32 整除铁律、单人发音锁定与全员闭嘴抑制指令装配、首帧 I2V 安全注入、分镜间显存防泄漏。

comfy-h3-video-producer

本地 ComfyUI 剧本转视频自动生产管线(MiniMax H3)。用户给剧本要求用 ComfyUI 出分镜视频,或提到"一采二采/超分/班长套件/CSGlide/工作流生成视频"时使用。覆盖:工作流握手(不写死节点号)、导演调度与段规划、H3 提示词、提交与自动监工、一采双检闸门、收片清床音与拼接、断点续产。

drama-audio-soundstage

短剧工业化管线·第 4 工序:声音工程与智能避让配乐。负责 H3 视频生成后的声音精修。硬核沉淀:谱减法音频门控(消灭风噪/电流床音)、Whisper ASR 纯普通话语种漂移拦截、FFmpeg 原生侧链动态避让混音(Sidechain Ducking,台词开口 BGM 瞬降 -14dB,闭嘴扬起)。

drama-director-staging

短剧工业化管线·第 1 工序:编剧与导演分镜调度。负责将剧情大纲或小说文本转化为符合短剧视听规律的工业级分镜。内置:6 大动态运镜轮换决策树、台词纯中文汉字归一化(数字大写)、短剧快节奏 20~28 字密度调控、单人发言与旁观者闭嘴标注。

drama-finishing-upscaler

短剧工业化管线·第 5 工序:终期成片与 1080p 超分增强。负责全集分镜无损拼接与院线级 1080p 超分辨率放大。硬核沉淀:FFmpeg concat 零损耗直拼、RealESRGAN x2plus GPU 独占 FP16 批量超分(~4.8fps)、原生高质量避让音轨无损混流、ffprobe 工业级物理参数验收。

Related Skills