Backblaze Launches Open-Source SDK to Streamline Generative Media Pipelines

Backblaze推出开源SDK以简化生成式媒体流水线 企业采用生成式媒体需要将碎片化的提供商API标准化，并利用事件驱动型存储来降低执行成本和延迟。 Jeronimo De Leon探讨了行业从以文本为中心的AI向以资产为中心的生成式媒体的转变，在此趋势下，模型会生成视频、高分辨率图像和交互式网页等复杂的数字资产。为了应对对接数十家不同模型提供商及其多样化延迟特性所带来的编排挑战，Backblaze开发了开源SDK GenBlaze，将非标准的API参数规范化为统一的执行基底。演讲重点阐述了事件驱动型云存储如何主动编排超分辨率放大与后期处理等工作流，从而在大规模应用中实现成本可预测性与高速执行。 Backblaze 应用 AI 总监 Jeronimo De Leon 指出，传统 S3 存储存在高昂且不可预测的出网费和 API 调用成本，而 Backblaze 目前为全球超 50 万家客户管理着 6 EB 的数据规模。 (0:04) 生成式 AI 正从文本输出转向以多媒体资产为中心，不仅能生成高分辨率音视频与可玩游戏，还能让玩具反斗城、可口可乐等主流品牌在 24 小时内完成以往耗时数周的热点营销内容制作。 (1:13) 面对 30 多家模型提供商主导地位频繁更迭的现状，企业级生成式媒体架构需要依靠流水线（Pipelines）、溯源（Provenance）和存储（Storage）三大核心支柱来避免系统重构。 (3:40) Backblaze 推出了被称为“生成式媒体领域 LangChain”的开源 SDK GenBlaze，用于统一各家模型的 API 参数，并利用 B2 存储的事件通知机制主动编排后续处理工作流。 (3:40)

Backblaze Director of Applied AI Jeronimo De Leon argues that generative AI is shifting from text outputs to asset-centric media production, necessitating unified pipeline abstractions and predictable cloud storage to handle fragmented provider APIs and infrastructure costs. The Evolution Toward Asset-Centric Generative Media (1:13) According to De Leon, generative AI is moving away from plain-text LLM outputs toward rich, multimodal digital assets, including high-resolution imagery, audio, video, interactive HTML dashboards, and playable games generated by tools such as Claude, Kimi 3, and Fable. This transition is accelerating turnaround times in commercial advertising, where brands like Toys "R" Us and Coca-Cola have incorporated generative AI into mainstream campaigns. De Leon points to real-time marketing during the NBA Finals—where a creator produced a viral Nike-style campaign featuring Victor Wembanyama and local nuns in under 24 hours—as proof that generative tools can compress multi-week production cycles into a single day. Normalizing Fragmented Pipelines With GenBlaze (3:40) Enterprise generative media requires chaining multi-stage workflows, spanning scripts, storyboards, audio design, video generation, upscaling, and analysis. However, developers must coordinate across more than 30 competing model providers whose performance leads shift rapidly between services like OpenAI, Runway, and Google. To mitigate integration friction caused by varying latency profiles and inconsistent parameters across media formats, Backblaze developed GenBlaze, an open-source SDK positioned as a LangChain-style normalization layer for generative media APIs. De Leon adds that cloud storage must function as an active orchestrator rather than a passive repository, using Backblaze B2 event notifications to trigger downstream processing, upscaling, and metadata extraction immediately upon asset upload instead of relying on continuous API polling. Infrastructure Scale and Predictable Storage Economics (0:04) Asset-heavy pipelines place heavy operational demands on underlying storage. De Leon contrasts Backblaze with Amazon S3, arguing that S3 customers routinely encounter unexpected ancillary charges, including high egress fees, API call costs, and regional replication billing. Backblaze, founded in 2007 and publicly traded under ticker symbol BLZ, manages six exabytes of data across 300 storage pods for more than 500,000 worldwide customers. De Leon emphasizes that transparent, low-cost storage architectures are vital for enterprise AI teams seeking cost predictability as media file sizes and generation volumes scale. As generative AI shifts toward asset-heavy multimodal workflows, development teams face significant friction from volatile model ecosystems and compounding cloud egress fees. De Leon contends that open abstraction layers and event-driven, cost-transparent storage are prerequisites for deploying generative media pipelines efficiently at enterprise scale.

Backblaze 应用 AI 总监 Jeronimo De Leon 指出，生成式人工智能正从文本交互加速迈向富媒体数字资产，企业需要通过标准化工具层与事件驱动的低成本存储来解决多模态管道的异构与碎片化挑战。 从文本走向多模态富媒体与即时营销 (1:13) De Leon 指出，生成式 AI 正从纯文本摘要转向包括音频、视频、高分辨率图像及交互式 HTML 页面在内的复杂数字资产。借助 Claude Co-work、Codeex、Kimi 3 与 Fable 等工具，系统已能依据提示词生成完整网页和可玩游戏。在商业应用中，玩具反斗城（Toys "R" Us）和可口可乐已将生成式 AI 纳入广告宣传活动；在 NBA 总决赛期间，创作者在 24 小时内即围绕马刺队修女与 Victor Wembanyama 制作出耐克风格的病毒式传播营销内容，大幅颠覆了传统数周的视频拍摄周期。 多模型管线集成与 GenBlaze 标准化 (3:40) 构建企业级媒体生成架构面临从剧本、分镜、视频生成、音频合成到超分辨率增强（upscaling）和后处理的多阶段复杂链条。目前开发者需要对接超过 30 家模型供应商，且市场主流模型每隔数周便会在谷歌、OpenAI 及 Runway 等厂商之间发生更迭。不同媒体格式的延迟特征和文件结构差异显著，为此 Backblaze 开发了开源工具 GenBlaze，定位为“生成式媒体领域的 LangChain”，旨在将各供应商非标准化的 API 参数统一抽象为标准执行基底。 事件驱动存储与云成本可预测性 (0:04) 在数据架构上，De Leon 主张存储不应是被动的存储桶，而应通过 Backblaze B2 的事件通知主动驱动工作流，在文件写入时自动触发后处理和元数据提取，免去轮询调用的开销。De Leon 还对比了 Amazon S3，指出其在存储本体费用外存在高昂的数据流出费（egress fees）、API 调用费及区域复制成本。成立于 2007 年的 Backblaze（美股代码 BLZ，年内股价上涨约 300%）目前服务超 50 万客户并管理 6 EB 数据，其核心定位是通过透明定价降低海量富媒体资产的存储负担。 生成式 AI 富媒体化促使多模态基础设施转向统一封装与主动编排存储，但面对频繁迭代的数十家上游模型接口，标准化架构能否持续跟进不同模态的性能与兼容需求仍有待观察。

## Background 本视频由 Ai4 频道于 2026 年 8 月 18 日发布在 YouTube 平台上，主讲人为 Backblaze 应用 AI 总监 Jeronimo De Leon。视频简介指出，随着生成式音视频正在重塑媒体产品架构，本次演讲面向管理层与架构师，系统性拆解了流水线归一化、C2PA 防篡改溯源以及零出网费对象存储三大技术支柱，为以资产为中心的 GenAI 时代提供工程落地指南。 ## TL;DR 演讲指出生成式 AI 正在由纯文本输出全面跨越至多模态的“资产中心”时代，端到端媒体生成已成为即时营销与数字资产生产的核心能力。面对市面上繁多的生成模型与极高迭代频率，企业需借助 GenBlaze 等 SDK 实现参数归一化，并以存储事件通知作为编排中枢，构建弹性的流水线；面对监管与版权合规，必须依托 C2PA 标准与 WORM（单写多读）不可变对象锁定技术，防止转码与分发过程中的血统元数据丢失；最终，在算力与模型接口频繁波动的背景下，企业需依靠自动化生命周期管理和低成本存储构筑起承载复杂计算的稳固“重力层”。 ## Atmosphere & style 整场演讲风格沉稳务实、逻辑清晰，演讲者凭借深厚的 AI 与 MLOps 实战背景，将抽象的架构设计与生动的工业案例（如文班亚马概念广告、可口可乐营销）巧妙结合。叙事节奏紧凑洗练，从模型迭代的工程痛点层层递进到防篡改存储与成本陷阱，直击架构师与团队决策者的核心诉求。整体氛围兼具技术前瞻性与极客实操质感，收尾对三大支柱的比喻铿锵有力，极具启发与说服力。 ## Key takeaways - [01:13](https://www.youtube.com/watch?v=8aqs0luPidg&t=73s) 生成式 AI 正在经历从文本交互向以资产为中心（Asset-centric）的范式转变，生成结果已演化为高清图像、4K 视频、交互式网页及游戏等数字媒体资产。 - [03:40](https://www.youtube.com/watch?v=8aqs0luPidg&t=220s) 生成式媒体系统由三大支柱构成：流水线（Pipelines）提供核心竞争护城河，溯源（Provenance）提供合规准入许可，而存储（Storage）则是维系一切的重力层。 - [03:40](https://www.youtube.com/watch?v=8aqs0luPidg&t=220s) 底层多模态模型迭代极快且接口异构，Backblaze 开源的 GenBlaze SDK 充当生成式媒体领域的 LangChain，实现跨供应商参数归一化与模型平滑回退。 - [03:40](https://www.youtube.com/watch?v=8aqs0luPidg&t=220s) 存储不应只是被动的终点站，Backblaze B2 的事件通知机制能够作为主动调度中枢，在上游生成文件落盘时自动触发后处理与超分流水线。 - [07:53](https://www.youtube.com/watch?v=8aqs0luPidg&t=473s) C2PA 规范正成为数字资产防伪与合规的事实标准，通过防篡改 JSON Manifest 记录提示词、模型版本及多轮迭代的分支血统与数字资产监管链。 - [11:10](https://www.youtube.com/watch?v=8aqs0luPidg&t=670s) 媒体元数据在转码、截图和社交平台分发中极易被剥离丢失，必须借助 WORM（单写多读）与对象锁定（Object Lock）在存储层保留不可篡改的出处凭证。 - [12:17](https://www.youtube.com/watch?v=8aqs0luPidg&t=737s) 第三方生成式 API 供应商普遍对生成内容设有过期销毁策略，企业必须主动拉取并转存自建存储，以保留原始数据用于后续模型微调。 - [12:17](https://www.youtube.com/watch?v=8aqs0luPidg&t=737s) 媒体生成过程中会堆积数倍于成品体积的草稿与中间分镜，必须在存储层配置自动化生命周期规则以避免存储成本失控。 - [14:08](https://www.youtube.com/watch?v=8aqs0luPidg&t=848s) 计算层与模型接口每隔数月就会面临重构与算力价格波动，而底层持久化存储具备长期稳定性与连续性，是抵御架构流动的核心定海神针。 - [14:55](https://www.youtube.com/watch?v=8aqs0luPidg&t=895s) Backblaze B2 通过零流出（Zero-Egress）费用、对象锁定与事件驱动编排，为端到端生成式媒体管道提供了低成本且高可靠的基座。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=8aqs0luPidg&t=4s) 演讲者背景与 Backblaze 的存储使命 演讲者 Jeronimo De Leon 拥有超过十年的人工智能行业经验。他的 AI 生涯始于 IBM 构建 Watson 项目，随后加入 MLOps 初创公司 Spell，期间参与推出了彭博社（Bloomberg）的首个对话式智能体；之后他在初创公司 Intellus 负责房地产数据的智能体工作流（Agentic Workflows）。目前，他担任开源云存储公司 Backblaze 的应用 AI 总监（Director of Applied AI）。Jeronimo 强调，Backblaze 的核心使命是让云存储变得“极致简单（Astonishingly easy）”，不仅在于使用体验，更在于透明简单的定价模式。他直言，许多使用 AWS S3 的团队不仅要承担基础存储费用，还会遭遇昂贵的出网流量费（Egress costs）、API 调用费以及跨区域复制费等隐性账单陷阱，而 Backblaze 则致力于消除这些复杂的成本开销。 - 成立于 2007 年，为纳斯达克上市公司（股票代码：BLZE），年内股价涨幅超 300%。 - 全球服务超过 50 万家客户，托管数据量突破 6 EB（Exabytes），由超过 30 万块机械硬盘（Spinning drives）构成的存储集群驱动。 --- ### [01:13](https://www.youtube.com/watch?v=8aqs0luPidg&t=73s) 从“以文本为中心”向“以资产为中心”的生成式跨越 生成式 AI 正在经历一场深刻的范式转移——从过去两三年令人惊叹但局限于文本的 LLM 输出，全面转向“以资产为中心（Asset-centric）”的多模态生成时代。无论是在办公协同工具中使用 Claude 生成交互面板与应用，还是利用 OpenAI、Codex 直接生成并托管 HTML 页面，亦或是使用 Kimi、Fable 等前沿模型直接从 Prompt 生成完整的游戏体验，AI 的终极产出已不再是干瘪的文字段落，而是包含报告、图片、高清视频、音频甚至沉浸式交互体验的多媒体数字资产。 这种实时资产生成能力正在重塑各行各业的营销与内容生产逻辑。Jeronimo 列举了广告传媒领域的标志性案例：玩具反斗城（Toys "R" Us）利用 GenAI 制作了整部商业广告片；可口可乐（Coca-Cola）将其大规模引入全球营销活动。尤为生动的是 NBA 总决赛期间的即时营销案例——马刺队比赛现场出现修女观众的画面曝光后，次日网络创作者便借助生成式工具，火速合成了维克托·文班亚马（Wemby）与修女同框的耐克风格概念广告并迅速引爆全网，品牌甚至无需组织实地拍摄便实现了对舆论热点的光速响应。 ```mermaid flowchart LR accTitle: 生成式 AI 资产形态演进 subgraph S1["文本时代 (Text-Centric)"] T1["Prompt 提示词"] --> T2["文本回复 / 摘要"] end subgraph S2["资产时代 (Asset-Centric)"] A1["多模态 Prompt"] --> A2["富媒体与代码"] A2 --> A3["图像 / 4K 视频 / 音频"] A2 --> A4["交互式 HTML / 动态面板"] A2 --> A5["端到端可玩游戏与交互体验"] end ``` --- ### [03:40](https://www.youtube.com/watch?v=8aqs0luPidg&t=220s) 生成式媒体流水线（Pipelines）与 GenBlaze 标准化方案 在深入剖析工程落地时，Jeronimo 将生成式媒体的架构体系划分为三大支柱：流水线（Pipelines）、溯源（Provenance）与存储（Storage）。传统自研模型的流水线通常包含数据摄取与冷热归档、数据预处理、模型实验与训练、部署推理及监控；而当前绝大多数媒体企业的核心工作流，则是依托海量基础大模型（Foundation Models），串联编排起包含“剧本编写、分镜脚本、视频生成、配音配乐、素材合成、超分辨率放大（Upscaling）与合规分析”在内的长链条复杂管线。 当前市场上活跃着超过 30 家媒体生成服务商，且底层模型正以“数周更新一次”的疯狂节奏迭代（从年初的 Google 图像模型到 OpenAI 与 Runway 的最新模型），不同提供商在接口参数、推理耗时和架构处理逻辑上差异巨大。为了避免企业因供应商更迭而陷入无休止的重构泥潭，Backblaze 推出了名为 **GenBlaze** 的开源 SDK。Jeronimo 将其形象地比喻为“面向生成式媒体模型的 LangChain”，旨在将异构的输入参数、模型调度与执行层全面归一化。 此外，Jeronimo 指出行业内常见的架构误区：“绝大多数人将存储视为一个被动的终点站，仅仅当成丢文件的地方。”他强调，存储应当是整个流水线的主动编排中枢。借助 Backblaze B2 的事件通知机制（Event Notifications），当上一阶段的生成文件落盘时，系统会自动触发后续的视频超分、元数据提取或分类任务，彻底告别了低效轮询服务状态的架构缺陷。 ```mermaid flowchart TD accTitle: 基于事件驱动与 GenBlaze 的媒体生成管线 subgraph Ingest["1. 输入与编排层"] P["多模态 Prompts / 脚本"] --> GB["GenBlaze SDK\n(参数归一化 / Fallback 链)"] end subgraph Gen["2. 多模态模型生成"] GB --> M1["图像生成模型\n(OpenAI / Google 等)"] GB --> M2["视频生成模型\n(Runway 等)"] GB --> M3["音频 / 旁白生成模型"] end subgraph Storage["3. 主动存储与事件中枢 (Backblaze B2)"] M1 & M2 & M3 --> B2["Backblaze B2 存储桶\n(落地即触发 Event Notification)"] B2 -.->|"自动触发下一阶段"| Proc["后处理流水线\n(合成 / 超分辨率 Upscaling / 元数据提取)"] end ``` --- ### [07:53](https://www.youtube.com/watch?v=8aqs0luPidg&t=473s) 资产溯源（Provenance）、C2PA 标准与防篡改链条 随着全球监管法规的密集出台，内容溯源与真实性验证已从合规建议演变成广播机构、保险公司及大型企业的强制性刚需。企业必须清晰自证：该资产由哪个模型版本生成？训练数据来源何处？迭代中间步骤为何？以彻底规避潜在的版权纠纷与法律风险。 针对图像真伪识别，单纯依靠隐藏水印（如 Google SynthID 等需要专用 AI 模型才能读取的机器学习隐形水印）仍无法提供完整的审计上下文。为此，由 Adobe 和微软等巨头推动的 **C2PA**（内容来源和真实性联盟）规范正在成为跨行业事实标准。C2PA 将全套断言（Assertions）封装为防篡改的 JSON Manifest 元数据文件，记录 Prompt 提示词、资产 Hash ID、父级运行记录（Parent run）、多轮迭代分支以及完整的数字资产监管链（Chain of custody）。 - **多轮迭代追踪**：区别于传统单次 Prompt 的纯文本对话，媒体生成往往需要经历多轮调优、重绘与合成，C2PA 清晰记录了从原始草稿到终稿的完整沿革。 - **只读防篡改（WORM）**：溯源元数据与媒体资产必须下沉并锁定在具备防篡改特性的存储层中，确保记录不可被非法篡改或静默删除。 - **开箱即用集成**：GenBlaze SDK 在执行多模型标准化的同时，内置了自动输出符合 C2PA 规范的 Provenance Manifest 功能，使开发者无需从零构建繁琐的合规审计管道。 ### [11:10](https://www.youtube.com/watch?v=8aqs0luPidg&t=670s) 出处追踪的脆弱性与不可变存储机制 在生成式媒体领域中，出处追踪（Provenance）与血缘管理（Lineage Tracking）正面临巨大的工程挑战。市面上虽然有大量生成式模型与媒体提供商，但这些第三方平台通常并不承担追踪与保留媒体生成血统的责任，这部分重担完全落在了构建媒体管道（Media Pipeline）的开发者肩上。开发者必须在存储层或文件本身内部建立完整的血统记录。 现实中，媒体的出处信息极易遭到破坏与剥离。最常见的断裂场景包括文件转码（Transcoding）、用户截图、剧烈的后期二次编辑，以及上传到各大社交媒体平台——绝大多数社交媒体在处理上传图片时都会主动抹除所有内置元数据。因此，必须将血统凭证剥离并独立保存在安全位置。解决这一问题的关键在于不可变存储（Immutable Storage），即 WORM（Write Once, Read Many，单写多读）存储策略。开发者可以为媒体生成对应的 JSON 出处清单（Provenance Manifest），并在对象存储层对其启用对象锁定（Object Lock），确保这些审计与血统记录无法被篡改或删除。 - **出处失效常见场景**：媒体转码、截图截取、社交平台元数据自动剥离、重度后期像素编辑。 - **WORM 防篡改保护**：在存储层将生成的 JSON 出处清单打上 Object Lock，确保证据链与审计合规性完好无损。 ```mermaid flowchart TD accTitle: GenAI Provenance Protection Flow subgraph Generation["生成与流转"] A["GenAI 媒体生成/编辑"] --> B["生成 JSON 出处清单"] A --> C["媒体分发/转码/社交上传"] C -. "元数据极易丢失".-> D["缺失 Metadata 的媒体"] end subgraph Storage["底层不可变存储 (Backblaze B2)"] B --> E["写入 WORM / Object Lock 存储"] E --> F["不可篡改的血统账本 (Lineage Ledger)"] D -. "关联校验".-> F end ``` ### [12:17](https://www.youtube.com/watch?v=8aqs0luPidg&t=737s) 资产爆炸、中间文件沉淀与生命周期管理 随着基础模型的迭代，模型输出的质量显著跃升，伴随而来的则是单文件体积与生成总量的几何级数增长。许多企业往往在遭遇严峻挑战时才仓促寻找替代方案——这通常是因为亚马逊 AWS S3 的账单金额已经失控，或是现有存储架构根本无法承载当前的数据吞吐。实际上，存储需求完全可以在管道设计初期进行前瞻性规划。 开发者不能依赖上游模型提供商来充当长久的数据归档库。绝大部分生成式 API 供应商都会在协议中明确标注：所有在线生成的内容都有严格的有效期限，过期即被彻底销毁。如果企业希望保留原始生成数据用于后续的模型微调（Fine-tuning）或构建自有原始数据归档（Raw Data Archive），就必须主动将所有生成内容拉取并保存在自己的存储架构中。 此外，用户往往误以为“最终导出的成品文件体积并不大”，却忽略了生成过程中堆积如山的中间子资产（Sub-assets）。从前期的多次 Prompt 尝试、草稿分镜到局部修复切片，海量的中间迭代版本会迅速占满存储空间。面对这种资产激增，系统必须在存储层配置自动化生命周期规则（Lifecycle Rules），自动判定何时降级或淘汰无用的中间临时文件。 - **提供商数据过期策略**：第三方 API 均设有输出保留期限，不具备归档属性，必须转存自建存储用于后续模型微调。 - **中间资产累积风险**：最终成品的体积只是冰山一角，数倍于成品的中间迭代版本需要生命周期规则自动清理。 - **存储成本规划**：避免在 S3 账单暴涨后被动应对，应提前规划大规模生成式媒体的高性价比存储方案。 ### [14:08](https://www.youtube.com/watch?v=8aqs0luPidg&t=848s) 响应式计算架构与作为“重力层”的持久存储 当前 AI 初创企业的技术栈呈现出高度的“响应式架构”（Reactive Architecture）特征。由于底层大模型每隔几个月就会推陈出新，许多团队平均每 6 个月就不得不重构一次系统管道。与此同时，计算资源（Compute）的市场价格也像“股票市场一样持续波动”，导致开发者调度算力的目标节点和模型接口总处于动态变化与频繁重评之中。 在这种瞬息万变的体系中，计算层是流动且脆弱的，而存储层则是更具长远价值的战略投资与定海神针。灵活多变的业务逻辑与计算编排可以轻盈地构筑在存储之上，无论前端计算架构如何因模型更迭而重构，底层存储都能保持极高的连续性与稳固性。 - **计算层的波动性**：模型迭代与 GPU 算力价格剧烈波动，导致计算管道每 3 到 6 个月面临一次重构。 - **存储层的长效价值**：存储作为持久中枢（Gravity），为上方灵活易变的响应式计算提供稳定支撑。 ### [14:55](https://www.youtube.com/watch?v=8aqs0luPidg&t=895s) GenBlaze SDK 与端到端媒体管道整合 一个标准的生成式媒体管道涵盖了从脚本编写（Script）、分镜图生成（Storyboards）、视频与音频合成（Video & Audio），到最终多轨混流剪辑（Composing）的全流程。为了简化这一复杂流程的工程落地，Backblaze 推出了开源的 `GenBlaze SDK`，并在 GitHub 上提供了名为 `GenBlaze media pipeline example` 的端到端开源范例，完整演示了从剧本撰写、分镜绘制到最终成片输出的自动化闭环。 ```mermaid flowchart LR accTitle: GenBlaze Media Pipeline Architecture subgraph Pipeline["GenBlaze 媒体处理流程"] S["脚本编写"] --> SB["分镜图生成"] SB --> AV["音视频合成"] AV --> C["多轨最终混流"] end subgraph B2Features["Backblaze B2 支撑特性"] E["事件通知 (Event Notifications)"] -. "驱动编排".-> Pipeline L["生命周期规则 (Lifecycle Rules)"] -. "中间件清理".-> AV O["对象锁定 (Object Lock)"] -. "出处防篡改".-> C end ``` Backblaze B2 贯穿了生成式媒体管道的每一个环节： 1. **输入源资产沉淀**：例如“图生视频”（Image-to-Video）中用户上传的高清原图输入。 2. **中间生成栈管理**：保存生成过程中的分镜与片段，并通过生命周期规则控制成本。 3. **事件通知驱动编排**：利用存储事件通知（Event Notifications）作为管道任务触发与工作流编排的基座。 4. **出处账本不可变锁定**：通过 Object Lock 保护 C2PA 与血统清单。 5. **低成本分发**：提供稳定的常态化存储成本与零流出（Zero-Egress）分发路径。 在演讲的尾声，Jeronimo De Leon 用一句话精辟总结了生成式 AI 媒体架构的三大基石：“管道是你的护城河，出处是你的运营准入许可，而存储则是维系一切的重力。”（"That pipeline is the moat, the provenance is the license to operate, and the storage is the gravity."） ## Notable quotes > [00:04](https://www.youtube.com/watch?v=8aqs0luPidg&t=4s) Astonishingly easy — Jeronimo De Leon > [01:13](https://www.youtube.com/watch?v=8aqs0luPidg&t=73s) 以资产为中心（Asset-centric） — Jeronimo De Leon > [03:40](https://www.youtube.com/watch?v=8aqs0luPidg&t=220s) 面向生成式媒体模型的 LangChain — Jeronimo De Leon > [03:40](https://www.youtube.com/watch?v=8aqs0luPidg&t=220s) 绝大多数人将存储视为一个被动的终点站，仅仅当成丢文件的地方。 — Jeronimo De Leon > [12:17](https://www.youtube.com/watch?v=8aqs0luPidg&t=737s) 最终导出的成品文件体积并不大 — Jeronimo De Leon > [14:08](https://www.youtube.com/watch?v=8aqs0luPidg&t=848s) 股票市场一样持续波动 — Jeronimo De Leon > [14:55](https://www.youtube.com/watch?v=8aqs0luPidg&t=895s) That pipeline is the moat, the provenance is the license to operate, and the storage is the gravity. — Jeronimo De Leon > [14:55](https://www.youtube.com/watch?v=8aqs0luPidg&t=895s) 管道是你的护城河，出处是你的运营准入许可，而存储则是维系一切的重力。 — Jeronimo De Leon ## Mentioned Jeronimo De Leon, Backblaze, Backblaze B2, GenBlaze, GenBlaze SDK, IBM, Watson, Spell, Bloomberg, Intellus, AWS S3, Claude, OpenAI, Codex, Kimi, Fable, Toys "R" Us, Coca-Cola, Victor Wembanyama, Nike, Google, Runway, Google SynthID, Adobe, Microsoft, C2PA ## Video description Jeronimo De Leon - Director, Applied AI, Backblaze Generative video, image, and audio are becoming core to a new generation of media products, and a clear architectural blueprint is emerging. The teams that turn this into a durable product capability will own the next decade of generative media. This keynote walks executives and architects through the three layers where the leaders are investing. Pipelines: the normalized job entity that absorbs new providers without rewrites, with parameter normalization, fallback chains, and event-driven storage as the orchestration substrate. Provenance: the trust contract that meets enterprise requirements at the moment regulators are codifying them, with C2PA, lineage, and chain-of-custody anchored to write-once-read-many storage. Storage: the long bet that compounds into a durable cost and capability advantage, with design tactics like content-addressable layouts, lifecycle automation, and zero-egress delivery paths that bend the cost curve by an order of magnitude at petabyte scale. A practical playbook for the asset-centric era of AI. [Source](https://www.youtube.com/watch?v=8aqs0luPidg)

[0:04] Jeronimo De Leon: Astonishingly easy [1:13] Jeronimo De Leon: 以资产为中心（Asset-centric） [3:40] Jeronimo De Leon: 面向生成式媒体模型的 LangChain [3:40] Jeronimo De Leon: 绝大多数人将存储视为一个被动的终点站，仅仅当成丢文件的地方。 [12:17] Jeronimo De Leon: 最终导出的成品文件体积并不大 [14:08] Jeronimo De Leon: 股票市场一样持续波动 [14:55] Jeronimo De Leon: That pipeline is the moat, the provenance is the license to operate, and the storage is the gravity. [14:55] Jeronimo De Leon: 管道是你的护城河，出处是你的运营准入许可，而存储则是维系一切的重力。

[0:04] Jeronimo De Leon: making storage astonishingly easy to use and affordably priced [3:40] Jeronimo De Leon: LangChain for generative media

[0:04] So a little bit about myself. I've been in  AI for the past 10 years. Initially starting [0:07] off uh at IBM building out IBM Watson projects.  Um then I moved to a ML ops startup called spell. [0:14] Rolled out Bloomberg's first conversational  agent and then more recently at a startup called [0:19] Intellus where we had agentic workflows for real  estate data. But now I'm at Back Blaze backlaz.com [0:25] the um open cloud storage company. And our mission  really is to make storage astonishingly easy. Uh [0:33] not only just to use but also from a pricing  perspective as well. Uh you probably if you [0:38] use currently S3 you not you don't only get get  uh hit with the storage costs but also so many [0:44] other costs as well such as um egress costs, API  core costs, uh region replication costs and more. [0:50] Um but we try to simplify that all that for you.  Uh we've been founded we were founded in 2007. [0:57] We're a public company BLZ actually uh probably  top performing company up 300% year to date. [1:03] Uh 500,000 customers worldwide six exabytes under  management and and using 300 spinning drives. [1:13] So this talk really is because we're going through  a movement now of from moving from text ccentric [1:18] kind of generative media to asset centric. um you  know the past couple of years everyone is really [1:24] kind of astonished of what LLMs can do but you're  you're starting to see more of those multimodal [1:29] generations and now the output is really asset  outputs if you're using kind of claude co-work [1:37] if you saw the latest Kimmy 3 models you're  starting to see the what's being produced or [1:43] what the output is are more kind of assets rather  than text uh and usually now you see even in the [1:50] work environment people are sharing dashboards  that they've built in co-work or codeex. You see [1:57] uh Codeex and OpenAI now having ability for you  to create a HTML output and then even host it on [2:03] their platform. And same with Claude uh as well.  So you're starting to see more and more outputs [2:09] being more assets with Kimmy 3 and actually the  latest Fable model. They're starting to produce [2:15] more even games, right? So the output from prompt  to whole experience a gaming experience and then [2:22] you'll probably see more experiential type  of outputs but you're that's the way you're [2:27] seeing these generative uh models uh move to where  they're not just producing text they're producing [2:34] either reports images video audio or uh HTML  outputs as well and you kind of look at we we want [2:43] to kind of look at the state of generative media  here where across different industries you're [2:49] starting to see more and more use especially kind  of in areas like advertising you saw Toys R Us [2:55] Coca-Cola in the past year actually using Genai  within their media campaigns uh Toys R Us created [3:02] a whole commercial uh using Genai this is just  an example of how reactive this was actually uh [3:09] created by a content creator but if you watch the  NBA finals there was some nuns at uh the Spurs [3:15] game and then the next day uh there was a whole  bunch pretty much a Nike media campaign that had [3:21] these WMY and the nuns in there. It was created  by a creator but then that campaign kind of went [3:27] viral for Nike as well. But you saw how fast they  were able to react. They didn't have to shoot a [3:32] campaign around that but they reacted because of  uh what happened in the media the the day before. [3:40] So we're going to talk about three core  things, pipelines, provenence, and storage. [3:44] And kind of how you need to be thinking about them  as you're kind of either building your generative [3:49] applications or building your generative  media pipelines as well. So Back Blaze, [3:54] we've been primarily like one of our core focuses  has been showing people how we kind of be embedded [4:00] within the model training pipeline and we've got  uh plenty of case studies that kind of highlight [4:05] that and have different type of products that  highlight that as well. But this is primarily [4:10] the high level generative media pipeline of the  data ingest and active archive the data processing [4:16] the model experimentation and training model  development and inference and then monitoring. [4:21] So you kind of bring it at a high level if you're  building your own models you've kind of go through [4:26] these stages as well and you kind of think through  all those different steps. But now with generative [4:30] media pipelines, there's more companies that are  primarily using those foundational models and [4:36] then thinking about what are all those different  steps to create primarily that final output. And [4:41] that's typically they're going through creating  scripts, creating storyboards, video, uh audio, [4:47] and then how do you compose that and upscale  and analyze that? And there's so many different [4:52] services that you can primarily use now uh through  to go through all those steps. there's probably [4:57] more than 30 providers uh and even more so now uh  especially presenting at the expo here that you [5:04] kind of had to think about or that you want to be  kind of thinking about and and using as well. Um [5:10] and a big reason of kind of thinking about  all those different type of riders is because [5:15] the models constantly evolve and update and  and you you want to be able to use the latest [5:21] and greatest. every it was previously every 6  months now every 3 months now every couple of [5:25] weeks there's a new generative model that has that  primarily is bestin-class uh remember at the start [5:32] of the year there was Google's nano Google's nano  banana uh with their generative image model but [5:38] now openai's generative image model kind of beats  that um and then you've got runway models that [5:43] are beating that as well so you constantly there  are constantly new providers coming out with new [5:49] higher better quality models so it's hard  to kind of just pick one and stick to it. [5:55] And this is where even as you kind of  think about that generative media pipeline, [6:00] you've got all these different assets you're  creating from images to video and audio and they [6:05] all have different types of architectures that  you have to be thinking about putting together. [6:09] Um, so when building your application, you  kind of handle different images differently [6:13] to how you handle video to how you handle  audio. and uh when they get generated the [6:19] time it takes to respond or the time it takes  to process those things are different as well. [6:24] So as you're kind of building out that pipeline,  what you want to be doing is thinking about how do [6:29] you standardize um all those models that you want  to be doing uh using all those inputs that that [6:35] you want to be taking in um and then also that  execution layer as well. And we've developed an [6:40] SDK called Gen Blaze that helps you experiment  across those different generative media models [6:45] primarily standardizing all those inputs  uh because they are uh so different and and [6:51] non-standardized across different models and  primarily uh placing that in one SDK for you [6:56] kind of tap into it. I like to think of it as  like lang chain for generative media models. [7:02] And then one thing to think about also as you're  kind of building out your pipeline is most people [7:06] think of storage as a passive destination where  this is storage is a place where I put my files. [7:12] Uh but sto storage primarily can be an active  part of your pipeline as well where you you [7:19] primarily use event notifications. Back blaze B2  has event notifications where as things land it [7:25] can trigger off the next step of your pipeline.  And that way your storage is actually the thing [7:29] that's helping files move along for it to be  processed to for it to be upscaled for it to be [7:35] have metadata uh extracted for it to be classified  because you you don't want to be sending [7:41] uh like polling the service to say is this file  ready is this file ready you primarily want to [7:45] have your storage play an active part in how  the kind of uh your pipeline works as well. [7:53] So talking about Providence, as you can see,  this is every year there's kind of more and [7:58] more happening within the Providence space  where there's new laws coming into place and [8:02] this is primarily to kind of start like putting  making people accountable of what those files [8:07] that have been generated are. Now you've  got even insurers or broadcasting having to [8:13] need that pro level of providence of how does  this content get generated? Where where was the [8:18] original models from? what where was those  source training files from? Because uh you [8:23] don't want to run into kind of copyright issues  later down the track after you've created a file. [8:29] And CTPA is one of the uh leading uh frameworks  that have kind of evolved from companies like [8:36] Microsoft and Adobe that have become the  standard of how do you kind of track uh [8:42] the uh generative media that's been generated  by by these models. It's still kind of an [8:47] evolving space as you can saw the timeline still  fairly recent but at least it provides a base of [8:52] claims and assertions of and then also gives  kind of more lineage trail of how this model was [8:58] created and also the steps that it takes created.  Unlike text where you primarily put one prompt in, [9:04] you get an output out. Usually these type of  media outputs have several steps of iteration [9:09] and you want to keep track of those  steps of iteration along the way. [9:13] Um, another way of kind of looking at providence  as well is like some people say why don't we just [9:18] place watermarking on the image and that is one  of the solutions as well. There's actually models [9:24] developed like this is an example from Google.  This is a Google synth ID model where you can't [9:28] actually even tell that there's a watermark within  that image and you actually need another model [9:33] to probably look at that image to kind of see  the watermark as well. So there's there's new uh [9:40] watermarking techniques that have been  developed and that's also MLbased to kind of [9:45] keep kind of the quality of the image and only  another AI would be able to tell it's AI generated [9:53] but water watermarking only provides you some of  the data and this is where that C2PA comes into [9:58] play where here is an example kind of starts  to list out the model that was created in the [10:03] version of the model and part of this is also  is eventually you want to potentially replicate [10:09] uh the output that you that you've  created and also want to see the steps [10:12] um the steps of the iterations  that that came across as well. [10:18] But as you can kind of see, Providence isn't  just one thing. It's kind of a stack of things [10:22] you kind of have to think about from that  CPA file that's al that's being created as [10:28] the final asset is being created the  watermarking that gets attached to it. So, [10:33] so that the CTPA file is primarily a JSON manifest  that has all those items such as prompt, hash ids, [10:41] the parent run, the iteration runs, uh,  and then also chain of custody as well. [10:46] Uh, where does that live? How's that being locked  down? So, it can't be deleted or modified as well. [10:53] And like you mentioned uh Gen Blaze the Gen  Blaze SDK that we have is also does that as [10:58] well. Besides providing you standardization  across uh the different generative meter models [11:04] uh it also generates outputs the only SDK that  generates that provenence output as well because [11:10] it's an area that we're thinking about and  totally understand that uh because we have a [11:14] handful or many uh generative media  providers and it's something that [11:19] they're they're not on the hook for. it's the  person that's developing that the media pipeline [11:24] that's on the hook of keeping the lineage and  tracking um within the storage or file itself. [11:31] But this is why we talked about having a  stack of the storage because there's so many [11:34] way that providence breaks uh from transcoding  to when people taking screenshots of the image [11:40] to social media uploads. When you upload an image  in social media uh it usually strips away a lot of [11:45] that metadata. So you have to have that lineage  kept somewhere else to heavy edits of the file. [11:51] And then mutable storage and this is  where um you know write once read many [11:57] storage comes into play where your storage layer  again plays a part with um you setting these JSON [12:05] manifest provenence files. you set object lock on  those areas so those files can't be tampered with [12:11] and you've got different uh features of uh  how you can lock lock them down as well. [12:17] So getting into the storage side of things,  storage is becoming more important because as [12:23] you can see as models kind of progress, file sizes  progress as well and that's something whether [12:28] your pipeline becomes more and more successful  you're going to have to be thinking about how [12:32] does the storage scale on top of that. It's funny  because people usually or companies come to us [12:38] when they realize actually um we've got a  really big problem now because the either the s [12:43] the S3 bill is really high or they they primarily  can't keep up with their current storage provider [12:48] so they end up coming to us. Um, but that's  something you can kind of plan ahead of time [12:53] because you you you know that as models  kind of progress that level of quality is [12:58] going to increase uh and those file sizes are  going to increase es especially if you're kind [13:03] of increasing uh the amount of items that you're  kind of generating within your pipeline as well. [13:08] And it's not up to those providers that you're  using to kind of be your uh storage archive. Like [13:14] they primarily state on their website that all  everything that gets generated by you will expire. [13:20] So it's up to you to kind of keep all everything  that you generate from those those providers [13:25] in some level of storage because you may want  to use them uh for fine-tuning or you want to [13:30] start building your raw data archive as well  from everything you've generated as well. [13:36] And this goes back to everybody thinks about  hey actually my final file isn't that big [13:41] but thinking about all those iterations of files  that uh led up to that final file is is something [13:47] to think that kind of adds up. And this is where  as you start generating more and more assets, [13:53] those those subassets start to build  up and you want to think about okay, [13:58] which ones do I discard? How do  I want to move them and start [14:02] uh moving them automatically with  life cycle rules within your storage. [14:08] One thing you start to think about um  and you kind of see a lot of startups, [14:12] they kind of start to work with reactive  architectures where they're rebuilding their [14:15] architecture every few months, every at least 6  months or so because either models are changing [14:22] um or their compute needs are changing as  well. Like you see computers like the stock [14:27] market these days where the compute prices  are fluctuating. So your compute is primarily [14:32] where you're pointing your compute or what what  models you're pointing to is constantly changing [14:36] uh or re being re-evaluated. But your  storage is more of that long-term play [14:42] uh and something that can be more fixed  uh because your architecture can primarily [14:47] sit on top of that storage and be flexible and  reactive while your storage is more consistent. [14:55] And this is where uh just wrapping it all  up. You've got that media pipeline of script, [15:00] storyboards, video, audio, and  then composing it all together. [15:04] We have that Gen Blaze SDK that kind of helps you  bridge that. And then the back blaze B2 storage [15:10] has all these features that kind of support  your generator pipeline from event notifications [15:15] for that will help with orchestration, object  lock for that providence private uh protection [15:20] and life cycle rules for that storage management.  And we've got a sample uh that's open source that [15:28] you can play with gen the Gen Blaze SDK. And  this is just an example where it kind of writes [15:32] the script, creates all the uh storyboard  images, then generates the video, audio, [15:37] and then final output as well. So if you search  for Gen Blaze media pipeline example on Google, [15:41] you'll run into this uh on GitHub. Um and  primarily as you can see back blaze B2 fits [15:50] across the entire stack from saving your inputs  because your inputs also have assets as well like [15:55] everyone usually says generate a video based off  this image. So you want to be saving all those [16:01] inputs especially if they're um photos and things  like that that people have been saving uh from [16:07] generative outputs uh intermediates or those  intermediates stack that we kind of showed [16:11] previously that provenence ledger distribution and  then that steady cost uh state cost as well. So we [16:18] we primarily have a place through throughout  all the different steps of your pipeline. [16:24] So just a final thought here where if you  think about that generative media pipeline [16:29] that pipeline is the mo the provenence is the  license to operate and the storage is the gravity [16:35] and we've got a booth so if you have  any questions or want to chat more [16:38] about how we could help you with a generative  media pipeline please come visit us. Thank you.