特别鸣谢
Special Thanks
感谢 NVIDIA 主办本次黑客松
NVIDIA — Hackathon Host
感谢 GPUS 开发者社区赛事支持
GPUS Community — Competition Support
感谢赞奇科技提供比赛设备支持
Zanqi Technology — Hardware Support
星 火 绘 卷
将经典文学长文本，经由六阶段多 Agent 工作流，
自动生成角色一致、印刷级品质的连环画页面。
Transform classic literature into illustrated comics through a
six-stage multi-agent AI pipeline — character-consistent, print-quality.
SparkScroll（星火绘卷）是面向经典文学内容生成的多模态创作平台。 围绕"长文本理解 → 剧情拆解 → 分镜生成 → 页面渲染 → 前端展示"构建完整闭环， 让《西游记》等经典名著化身精美的数字连环画。
SparkScroll is a multimodal creation platform for classic literary content. It closes the loop from "long-text understanding → plot decomposition → storyboarding → page rendering → frontend display," transforming masterpieces like Journey to the West into beautiful digital comics.
项目名称中，Spark 致敬 NVIDIA DGX Spark 平台， Scroll 代表连环画与画卷； 中文"星火绘卷"兼具科技感与古典文学韵味。
In the name: Spark pays tribute to the NVIDIA DGX Spark platform, and Scroll represents comics and picture scrolls. The Chinese name "星火绘卷" carries both a sense of technology and classical literary charm.
Director → Writer → CharacterDesigner → Drafting → Editor → Assembler， 完整还原人类漫画制作流程，从百万字长文到精美页面一气呵成。
Fully reproduces the human comic production workflow — from million-word novels to polished pages in one continuous pipeline.
立绘 Agent 预生成角色正/侧/背三视图与道具参考图，后续所有分镜以参考图为基准注入扩散模型条件通道，风格统一、叙事连贯。
CharacterDesigner pre-generates front/side/back reference sheets; all subsequent panels use these as conditioning inputs to the diffusion model, ensuring visual consistency throughout.
文本主脑（~40 GB）与视觉中枢（50–60 GB）同时常驻 DGX Spark 128 GB 统一内存，零 Swap、极致秒级响应，多 Agent 极速切换。
Text brain (~40 GB) and visual core (50–60 GB) co-reside in DGX Spark's 128 GB unified memory — zero swap, second-level response, rapid multi-agent switching.
Editor Agent 使用 Python PIL 在像素层面排版文字，彻底消除 AI 生字乱码，保证中文字体清晰、布局精准、达到出版级品质。
The Editor agent uses Python PIL to typeset text at the pixel level — no AI-generated garbled characters, ensuring crisp Chinese fonts and publication-grade layout quality.
基于 128K 上下文对长文本进行"降维压缩"，提取核心剧情节点与角色关系，输出结构化剧本大纲。
Leverages 128K context to compress long texts, extracting core plot nodes and character relationships into a structured story outline.
将大纲拆解为分集、分镜，每页 4–6 个场景，输出标准化 JSON 剧本，指定每格台词与画面描述。
Breaks the outline into episodes and panels (4–6 scenes per page), outputting a standardized JSON script with dialogue and visual descriptions for each cell.
预生成每个角色的正面、侧面、背面三视图和关键道具参考图，为后续分镜锁定视觉一致性基准。
Pre-generates front/side/back three-view character sheets and key prop reference images, establishing the visual consistency baseline for all subsequent panels.
基于分镜 JSON 和角色参考图，调用 Diffusers + vLLM-Omni 双驱动生成各格底图，保证角色风格与参考图对齐。
Using the storyboard JSON and character references, calls the Diffusers + vLLM-Omni dual engine to generate panel base images, keeping character style aligned to references.
纯 PIL 像素级排版：将台词嵌入对话框，精准控制字体大小、行距与位置，拒绝 AI 乱码，输出印刷级清晰图像。
Pure PIL pixel-level typesetting: embeds dialogue into speech bubbles with precise font control, eliminating AI-generated garbled text and producing print-quality images.
聚合各页产物，流式按页推送给前端，每生成一页即时呈现，用户无需等待全集完成即可阅览。
Aggregates per-page outputs and streams them to the frontend in real time — users can start reading as soon as the first page is ready, without waiting for the full episode.
SparkScroll 采用"重型双大模型常驻显存"架构，普通单卡系统完全无法运行该并发管线。
SparkScroll's "Memory-Resident Multi-Model" architecture simply cannot run on ordinary single-GPU systems.
| 模块 | Module | 选型 | Selection | 显存预算 | Memory Budget | DGX Spark 必要性 | Why DGX Spark |
|---|---|---|---|---|---|---|---|
| 文本与逻辑主脑 | Text & Logic Brain | Qwen3.5-9B 128K ctx · vLLM v0.19.0 |
~40 GB | ~40 GB | 128K 上下文足以覆盖绝大多数小说全文，稳定支撑剧情压缩与跨集规划。 | 128K context covers most full-length novels, enabling stable plot compression and cross-episode planning. | |
| 视觉渲染中枢 | Visual Rendering Core | Qwen-Image-Edit-2511 FireRed-Image-Edit-1.1 |
50–60 GB | 50–60 GB | 兼顾高保真角色一致性编辑与中文字幕渲染，强行量化会导致排版几何推理能力丧失。 | Balances high-fidelity character editing and Chinese subtitle rendering; forced quantization destroys layout geometry reasoning. | |
| 框架与并发缓冲 | Framework & Concurrency | vLLM · FastAPI · Diffusers · OS | ~15 GB | ~15 GB | 支撑 API 流转、自研调度与多 Agent 极速切换，需安全冗余。 | Supports API routing, custom scheduling, and rapid multi-agent switching, requiring safe redundancy. | |
| 总计 | Total | — | 110–115 GB / 128 G | 110–115 GB / 128 G | 只有 DGX Spark 能实现"零 Swap"的极致秒级响应，支撑多 Agent 极速切换与流式呈现。 | Only DGX Spark achieves zero-swap, second-level responses enabling rapid multi-agent switching and streaming delivery. |
选自《西游记》《卖火柴的小女孩》《小红帽》等经典名著
Samples from Journey to the West, The Little Match Girl, Little Red Riding Hood, and more
将《西游记》等古典名著转化为连环画，提升学生知识留存率与阅读兴趣。
Transform classics like Journey to the West into comics, boosting student retention and reading engagement.
网络小说每日更新章节自动生成 5–10 页连环画，在短视频平台快速引流。
Auto-generate 5–10 comic pages per chapter update, driving traffic on short-video platforms.
将晦涩的历史文献转化为连环长卷，让博物馆展览更加生动易读。
Convert obscure historical texts into scrolling comic murals, bringing museum exhibitions to life.
品牌故事快速转化为多风格连环画，支持多语言适配，降低内容本土化成本。
Rapidly convert brand stories into multi-style comics with multilingual adaptation, reducing localization costs.
重建年印量曾达 81 亿册的连环画产业，让"小人书"以数字化形态涅槃重生。
Revive the comic-book industry that once printed 8.1 billion copies annually — reborn in digital form.
团队起名"纵贯线"，源自成员来自北京、南京、广州三地，如一条纵贯全国的线。 三地协作、远程共享 DGX Spark 设备，以 OpenClaw 多 Agent 辅助开发。
"Zongguanxian" means the line running through the country — a nod to the team spanning Beijing, Nanjing, and Guangzhou. Remote collaboration, shared DGX Spark access, and AI-assisted development via OpenClaw.
队长 · 项目策划
Team Lead · Project Planning
项目策划与管理、本地模型部署、前端 UI 开发与调试
Project planning & management, local model deployment, frontend UI development
队员 · 架构设计 · 代码负责
Architecture Lead · Code Owner
系统整体架构、技术路线制定、Gateway 开发负责人
System architecture, technical roadmap, Gateway development leader
队员 · 测试与文档
Testing & Documentation
前端 UI 开发、系统测试、技术文档、参赛材料编写
Frontend UI, system testing, technical documentation, competition materials
AI 助手 · Gateway 开发
AI Assistant · Gateway Dev
设计和实现后端 API 服务
Designed and implemented backend API services
AI 助手 · 前端 UI 开发
AI Assistant · Frontend Dev
设计用户友好的前端界面，提供直观的操作入口
Designed user-friendly frontend interfaces with intuitive model operation flows
感谢 NVIDIA 主办本次黑客松
NVIDIA — Hackathon Host
感谢 GPUS 开发者社区赛事支持
GPUS Community — Competition Support
感谢赞奇科技提供比赛设备支持
Zanqi Technology — Hardware Support