这期播客发布于2026年9月18日，Gorkum 和 Batuan 在其中讨论了 FALS 突破性的视频生成模型 H3 Max 及其对生成式媒体领域的影响。他们将 H3 Max 介绍为在速度、质量和成本效益方面“独树一帜”，并指出基准测试常常需要对数刻度才能显示其与其他模型的区别。 H3 Max 的灵感来自于 Minimax 的 H3 模型，H3 模型是第一个真正开源的、最新一代的视频模型，能够接收参考输入。FALS 作为一个通用的媒体推理服务平台，看到了一个“全力投入”对该模型进行后训练的机会，旨在实现显著的改进。他们将市场机会定义为“代币市场契合”，即个人能够卓有成效地花费大量代币（例如数千美元）来完成日常视频生成任务。该行业一直受限于计算资源，而 H3 Max 提供了一个“效率提高一个数量级”的解决方案，释放了计算资源并使得能够生成更多的代币。 Batuan 详细阐述了技术进步，指出虽然系统范围的优化通常能使模型提速 2-3 倍，但 H3 Max 在此基础上又实现了“一个数量级”的提升。这是“系统/模型协同设计”的结果，其中包括对模型进行后训练，使其在更少的步骤（例如 20 步而不是 50 步）下运行，同时保持甚至提高质量。此外，FALS 实施了广泛的内核和系统工程，将硬件利用率从通常的 30-40% 提升到理论最大 FLOPS 的 70-80%。整个管道，从使用大型语言模型进行提示扩展到扩散、VAE 解码和超分辨率，都得到了专门的优化。这种复合效应使得相比原始的 Minimax H3 端点实现了“35 倍的速度提升”和“数量级更低的成本”。尽管 GB200 等新硬件提供了 2-3 倍的速度提升，但成本仍然相当，这凸显了 FALS 的优化超越了纯粹的硬件能力。 他们还发布了 H3 Max Turbo，这是一个公共版本，可以在大约 1.5 秒内生成一段五秒钟的视频，成本降低 2 倍，同时保持 97% 的质量。发言者指出，当前的速度和成本已经优化到极致，因此目前的重点是提高质量和可控性。 H3 Max 的发布在 FALS 内部激发了巨大的创造力。一名工程师 Rehan 自发地在 Twitch 上从他的笔记本电脑上连续直播 H3 Max 的生成内容。与此同时，Twitter 上的网红 Levels IO 推出了一个类似的无限流媒体网站。在内部，FALS 机器学习团队开发了 H3 Max 的连续版本，名为 H3 Max Director，它具有“两分钟的记忆”和“不断演进的系统提示”，以实现长达 60 分钟的连贯性。这使得实时、动作控制的视频生成成为可能，例如，一个新角色可以进入场景，而办公室和现有角色保持一致。FALS 通过其“Fall Live”网站展示了这一点，该网站迅速走红，允许众包控制视频的进程。 H3 Max 已迅速成为 FALS 平台上最受欢迎的视频模型，其使用量是其他模型的两倍。发言者强调了它对好莱坞和专业创作者的潜力。该模型的速度支持一种流行的工作流程，即艺术家使用 Blender 渲染低分辨率场景，然后将其作为参考输入到 H3 Max，以实现近乎 100% 的可控性。这解锁了新的管道，尤其当与 GPT Astra 等大型语言模型结合使用时。 FALS 现在正大力专注于增强可控性。即将推出的功能包括唇形同步（将音频与角色嘴唇匹配）、动作控制（将舞蹈动作应用于 AI 角色）和精确的摄像机控制（用于在不同时间点指定摄像机角度和移动的基于 JSON 的输入）。这些功能建立在 FALS 的后训练基础设施之上，该基础设施允许他们根据特定的专业需求调整任何模型，解决了研究与创作者需求之间的“脱节”。这包括为工作室进行知识产权整合和美国托管数据驻留（适用于 C-US 模型）。即将召开的 GenMedia 大会反映了这一转变，与会者现在以好莱坞工作室和新的 AI 工作室为主，这预示着 AI 在娱乐行业的应用将增长 10 到 100 倍。

This podcast, published on 2026-09-18, features Gorkum and Batuan discussing FALS's groundbreaking video generation model, H3 Max, and its impact on the generative media landscape. They introduced H3 Max as being "in a league of its own" for its speed, quality, and cost-effectiveness, noting that benchmarks often require a log scale to show its separation from other models. The inspiration for H3 Max came from Minimax's H3 model, which was the first truly open-source, latest-generation video model capable of taking references. FALS, working as a general media inference serving platform, saw an opportunity to "go all in" on post-training this model, aiming for significant improvements. They define the market opportunity as "token market fit" where individuals productively spend large amounts of tokens (e.g., thousands of dollars) for daily video generation tasks. The industry has been compute-constrained, and H3 Max offers an "order of magnitude more efficient" solution, freeing up compute and enabling more token generation. Batuan elaborated on the technical advancements, stating that while system-wide optimizations typically make models 2-3x faster, H3 Max achieved an "order of magnitude" beyond that. This was a result of a "system-slash-model co-design" involving post-training the model to run on fewer steps (e.g., 20 instead of 50) while maintaining or even improving quality. Additionally, FALS implemented extensive kernel and systems engineering, boosting hardware utilization from a typical 30-40% to 70-80% of theoretical maximum FLOPS. The entire pipeline, from prompt expansion using LLMs to diffusion, VAE decoding, and upscaling, received dedicated optimizations. This compound effect led to a "35x speed-up" compared to the original Minimax H3 endpoint and "magnitude lower cost." While new hardware like GB200 offers 2-3x speed improvements, the cost remains comparable, highlighting that FALS's optimizations go beyond raw hardware power. They also revealed H3 Max Turbo, a public version that generates a five-second video in about 1.5 seconds, costing 2x less while maintaining 97th percentile quality. The speakers noted that the current speed and cost are so optimized that the immediate focus is now on improving quality and controllability. The release of H3 Max sparked immense creativity within FALS. An engineer, Rehan, spontaneously started streaming continuous H3 Max generations on Twitch from his laptop. In parallel, a Twitter influencer, Levels IO, launched a similar infinite streaming website. Internally, the FALS ML team developed a continuous version of H3 Max, called H3 Max Director, with a "two minutes of memory" and an "evolving system prompt" for coherence up to 60 minutes. This allows for real-time, action-controlled video generation where, for example, a new character can enter a scene, and the office and existing characters remain consistent. FALS demonstrated this with their "Fall Live" website, which went viral, allowing crowdsourced control over the video's progression. H3 Max has rapidly become the most popular video model on the FALS platform, doubling the volume of other models. The speakers emphasized its potential for Hollywood and professional creators. The model’s speed enables a popular workflow where artists use Blender to render low-resolution scenes, then feed them as references to H3 Max for nearly 100% controllability. This unlocks new pipelines, especially when combined with LLMs like GPT Astra. FALS is now heavily focused on enhancing controllability. Upcoming features include lip synchronization (matching audio to character lips), motion controls (applying dance movements to AI characters), and precise camera controls (JSON-based input for specific camera angles and movements at different time points). These features are built on FALS's post-training infrastructure, which allows them to adapt any model to specific professional needs, addressing the "disconnect" between research and creator requirements. This includes IP integration for studios and US-hosted data residency (for C-US models). The upcoming GenMedia conference reflects this shift, with attendance now dominated by Hollywood studios and new AI studios, signaling a 10x-100x increase in AI usage in the entertainment industry.