Higgsfield Zephyr: Special 全开源：K-pop 机甲 AI 电影的完整制作工艺

by Mycelium Protocol

---

这是 AI 视频领域第一次有人把一部机甲 K-pop 短片的完整制作过程拆开来给你看：提示词、角色参考图、失败镜头、成功镜头，连创作过程中走过的弯路都留着，全部公开。

《ZEPHYR: Special》，Higgsfield 出品，约 5 分 7 秒，使用 Seedance 2.5 制作，发布后 16 小时内超过 11,000 次播放。

完整项目文件：https://higgsfield.ai/original-series/zephyr-special/full-film
（在 Cinema Studio 中打开即可查看所有提示词和资产）

---

这是谁的故事

《Zephyr》的世界里，有一支由五个女孩组成的小队——她们既是 K-pop 偶像，也是驾驶机甲的战斗精英。五个角色，每人都有独立的性格设定、服装设计、专属机甲：

• MIRA — 战略型领袖，机甲配备火焰喷射器
• REINA — 近战专家，爆发力强
• NAOMI — 技术型成员，机甲装备钩爪，是《Special》这集的核心视角
• ZERO — 冷静型狙击手
• HARUMIN — 支援型角色，小队的情感纽带

《Special》的剧情：Naomi 深陷险境，机甲倒扣，生死悬于一线。就在此时，一位神秘的客座角色出现，用超越小队认知的机甲战斗技巧将她解救。

---

两个「不可能镜头」

主创团队在开源说明里重点披露了两个技术难题，以及他们最终如何解决它的。

技巧 1：倒置角色参考图

Naomi 的机甲在这场戏里是倒扣着的。他们需要画面传达出失重感和混乱感。

第一个方案：精确提示词描述物理状态。

失败。模型生成的结果没有准确捕捉倒置状态下的物理感。

解决方案：把问题推到输入阶段。

他们把角色参考图直接倒置——把物理状态「烘焙」到输入里，而不是试图在提示词里解释物理原理。这样，提示词就不需要再解释「为什么是这个姿态」，可以专注描述画面的细节和情绪。

结果：Naomi 有了一个专属的「倒置状态」角色参考图。

工程层面的核心假设：模型更擅长从视觉输入推断物理状态，而不是从文本描述重建物理状态。 把物理信息编码进图像输入，是绕开文本-物理转换损耗最有效的方式。

---

技巧 2：锚点式最小提示词

Naomi 被从机甲里救出来的镜头，是这集里最紧张的画面之一：需要同时传达危险感、机甲的巨大重量、Naomi 的体力和无畏，以及她骨子里的优雅和脆弱。

他们没有用精确描述每一帧动作的方式写提示词。

相反，他们只设定了几个关键锚点，把更多创作自由留给模型：

我们想传达的是危险感、机甲的巨大重量、Naomi 的体力和无畏——在不失去她的优雅和脆弱的前提下。

结果：模型在锚点约束内自主填充了动作细节，反而产生了他们原本想要但无法用精确提示词复现的那种「真实感」。

工程层面的核心假设：复杂动作场景中，过度精确的提示词会压缩模型的创作空间，导致僵硬感。给定结果期望（情绪/物理结果）而非过程描述，能得到更自然的动作。

---

开源了什么

《ZEPHYR: Special》一共公开了 197 个资产，包括：

• 所有使用的提示词（无一保留）
• 最终入选剪辑的镜头
• 失败镜头（bloopers）——包括走过的弯路和被淘汰的方案
• 角色参考图（包括「倒置状态」等特殊状态版本）

任何人都可以在 Higgsfield Cinema Studio 里打开这个项目，看到完整的生成记录，从第一版到最终版本，一步步的迭代过程完全透明。

---

第一集的工程教训（Seedance 2.0）

团队同样公开了《Zephyr》第一集（使用 Seedance 2.0 制作）的制作经验，那时他们还在摸索这条生产流水线。

机甲角色参考图：细节的双刃剑

最初，他们制作了包含所有细节的「主参考图」：武器系统、舱盖开合机制、各种状态。

问题：模型看到细节，就会尝试展示细节。

具体来说：机甲的舱口在角色参考图里是开着的（作为一种状态说明），结果模型几乎在所有场景里都把舱口画成开着的，或者在「应该关闭」和「开着」两个状态之间形成冲突，导致机甲外形发生变形。

解决方案：

1. 多版角色参考图，分状态管理：把武器收纳状态和展开状态分成两张参考图。只有场景里需要展示武器时，才上传带武器的参考图。

2. 局部特写作为独立输入：如果需要展示武器的具体工作方式（比如 Mira 的喷火器怎么伸出来），把这个部件的特写图单独上传，而不是期望模型从整体参考图里「聚焦」到这个细节。

3. 参考图上的文字说明几乎无效：「Retractable revolver in the arm」之类的注解对模型没有实质帮助——你还是需要在提示词里完整描述武器的动作过程。别依赖参考图上的标注。

样式前缀的重要性

《第一集》里，他们还没有建立统一的样式前缀（style prefix）。这导致不同场景之间的视觉风格不够统一。

建议：在开始制作前建立一个固定的样式前缀（打光风格、色调、摄影机语言），放在所有提示词开头，保持整集的视觉一致性。

---

角色创建流水线

整条角色创建流水线分三步：

第一步：角色设定

用 Higgsfield Supercomputer 或 Claude 进行角色性格头脑风暴。主创团队明确说：「你可以很快通过与 AI 协作，获得那层基础深度——别让知识不足成为创作的障碍，直接开始吧。」

性格设定的作用不只是故事，更是生成指南：知道角色的性格，才能准确描述她的表演、反应和潜在故事弧。

第二步：服装和形象设计

用 Higgsfield Soul 2 进行角色和服装创作。Soul 2 的时尚能力几乎没有限制，可以处理任意风格混搭和细节组合，用来设计偶像 × 战士的复合形象效果最佳。

第三步：角色参考图精修

用 Nano Banana 2 Pro 和 Seedream 把设计稿整合成最终角色参考图资产。

---

Seedance 2.5 改变了什么

《Zephyr》第一集用的是 Seedance 2.0（那时刚发布，团队还在大量实验）。

《Zephyr: Special》用的是 Seedance 2.5。主创团队在项目说明里直接说：

今天，AI 创作者在与「演员」合作方面的体验，已经越来越接近在真实片场的感受。技术层面的取巧和与模型的博弈正在退场，这让你能更深入地沉浸到创作过程中，打磨那些细微之处。

Seedance 2.5 带来的最关键变化：角色一致性和场景控制能力大幅提升——这是那两个「不可能镜头」能在没有妥协的情况下实现的底层原因。

---

Higgsfield 平台生态

Higgsfield 不只是一个视频生成工具，它在构建一个AI 原生的内容生产平台：

工具	用途
Cinema Studio	场景控制、项目管理、生成历史追踪
Soul 2	角色/人物/服装创作
Nano Banana 2 Pro	图像精修
Seedream	角色参考图整合
Seedance 2.5	主视频生成模型（ByteDance）
Supercomputer	AI 辅助头脑风暴
Academy	制作流程课程（免费）

Higgsfield Academy 目前有几门重点课程与这个话题直接相关：

• Blockbuster 4K: The AI Filmmaking Pipeline（40 分钟，从脚本到逐场景提示词工程）
• Build an Ultra-Realistic Short Film in 4K（33 分钟，Claude Fable 5 + Seedance 2.0 4K 足球剧）
• Add AI VFX to Real Footage（12 分钟，真实素材 + AI VFX 合成）

Higgsfield Global Film Festival 目前正在进行，奖金池 100 万美元。Higgsfield 选择在此时开放《Zephyr》和《Hell Grind》的完整制作文件，意图很明显：帮助参赛者真正理解生产工艺。

---

《Hell Grind》：另一个开源参照

同期，Higgsfield 也开源了另一部原创系列《Hell Grind》第一集——四个街头少年意外得到神秘力量。这集的规模更大：

• 365,561 次观看
• 115,446 次生成（公开的完整生成历史）
• 3,138 个公开资产

这两套开源素材合在一起，构成了目前 AI 视频领域最完整、最透明的生产案例库之一。

---

实战总结

从《Zephyr》整个开源项目里，可以提炼出一套 AI 电影工业的核心工程思路：

1. 把物理状态编进输入，别试图用文字解释物理 ——倒置参考图技巧
2. 复杂动作场景：只给锚点，不给过程 ——锚点式最小提示词
3. 多状态角色参考图 ——每个重要状态（收纳/展开/特殊）单独一张，别共用
4. 局部特写优先于整体参考图 ——细节展示用独立输入
5. 样式前缀统一视觉语言 ——在所有提示词开头建立一致的打光/色调/摄影机前缀
6. 先做性格，再做生成 ——知道角色是谁，才能准确描述她该怎么演
7. 公开失败镜头是最好的工程文档 ——bloopers 比成功镜头更能说明问题在哪里

---

手把手教程：个人 / 小团队如何用这套方案做出科幻大片

前提：你有 Higgsfield 账号（可免费注册），能访问 Seedance 2.5。建议先打开《Zephyr》项目文件对照着看：https://higgsfield.ai/original-series/zephyr-special/full-film

---

阶段一：前期开发（Pre-production）—— 2-3 天

这是整条流水线里投入时间最值得的阶段。跳过它会让你在生产阶段付出双倍的时间成本。

第一步：用 Claude 或 Higgsfield Supercomputer 建立世界观和角色

把以下这个模板粘贴给 Claude 开始头脑风暴：

我想制作一部约 3-5 分钟的科幻动画短片。

核心设定：[一句话描述你的世界，例如「2087 年废土城市，人类和机械生命体共同生活」]
主角数量：[1-3 个，个人制作控制在 2 人以内]
核心冲突：[一句话，例如「失忆的雇佣兵发现自己其实是被设计出来的武器」]

请帮我为每个角色生成：
1. 外形关键词（3-5 个，便于后续生图）
2. 性格词条（3-5 个，用于写表演描述）
3. 标志性动作或习惯（1-2 个，用于区分角色的镜头语言）
4. 与其他角色的关系张力

为什么要先做这步：Zephyr 团队明确说了——「知道你的角色，才能准确写出他们的表演、反应和故事弧」。角色设定文档是后续所有提示词的参考依据，而不只是故事背景。

第二步：定义「样式前缀」

在生成任何镜头之前，先确定你的视觉语言。把以下要素写成一段固定文字，放在后续所有提示词的开头：

[样式前缀示例]
Cinematic, 4K, anamorphic lens, neon-noir lighting, cool blue and amber color grade, 
shallow depth of field, film grain, [你的世界名]

Zephyr 第一集没有固定样式前缀，导致跨场景视觉不统一。这是他们明确点出的错误。你的样式前缀就是整部片的「美术圣经」。

---

阶段二：角色资产制作 —— 2-5 天

第三步：用 Higgsfield Soul 2 生成角色形象

Soul 2 是专门做角色和服装设计的。打开 Higgsfield → Image → Soul 2。

每个角色分两轮生成：

第一轮：找对外形

提示词格式：
[性别] [发色] [发型], [外形特征 3 个], [服装风格], 
[配件或标志性元素], sci-fi [你的风格词], full body, 
white background, character design sheet

第二轮：固定「英雄形象」
选最好的一张，继续在 Soul 2 里迭代，直到你认为「这就是他」。这张图是后续所有生成的锚点，不要随意改动。

第四步：制作多状态角色参考图

这是 Zephyr 团队踩坑最多的地方。每个角色至少需要以下几张参考图：

参考图类型	用途	备注
标准站立（正面 + 侧面）	所有常规场景	主力参考图
情绪特写（2-3 种表情）	对话/情绪场景	只需上半身
动作姿态	战斗/奔跑场景	不要包含武器细节
武器/道具展开（单独一张）	使用武器的场景	只在需要时上传

用 Nano Banana 2 Pro 精修每张参考图，让细节更清晰稳定。用 Seedream 把多张设计稿合并成统一风格的参考图组。

关键：武器和机关永远不要出现在「标准站立」参考图上。 模型看到就会渲染。

第五步：科幻载具 / 机甲设计（如果有）

机甲比人形角色需要更多状态管理：

必须单独准备的参考图：

1. 全身标准形态（关闭所有特殊部件）
2. 驾驶舱区域特写（舱盖关闭状态）
3. 驾驶舱区域特写（舱盖打开状态）
4. 武器系统特写（每种武器单独一张）
5. 倒置/受损状态（如果剧情需要）——参考 Zephyr 的倒置参考图技巧

注意：凡是会开合、伸缩、变形的部件，都需要「收起状态」和「展开状态」两张独立参考图。

---

阶段三：生产前测试 —— 半天

在正式制作之前，用你的参考图做一轮「压力测试」，验证每个资产的稳定性。

第六步：资产稳定性测试

在 Cinema Studio 里，用以下模板对每个角色参考图生成 3-5 个测试镜头：

[样式前缀]

[角色名], [外形关键词 2-3 个], standing in [中性环境], 
looking at camera, medium shot, cinematic.

Camera: static shot, locked tripod.

检查这 3-5 个镜头：

• 角色外形是否稳定（发色、服装、面部特征）
• 有没有出现意外的细节（武器、机关、不该有的元素）
• 光照和色调是否与你的样式前缀匹配

如果不稳定 → 回到第四步调整参考图，再测试。通过了才进正片生产。

---

阶段四：正片生成 —— 一场戏一场戏来

第七步：把脚本拆成「镜头清单」

不要按场次想问题，要按镜头想问题。每个镜头清单条目包含：

镜头 [编号]
- 人物：[谁]
- 动作：[做什么]
- 情绪：[什么感觉]
- 摄影机：[机位 + 运动]
- 时长：[3-8 秒]
- 参考图：[用哪张]
- 特殊技巧：[倒置/锚点/等]

一个 5 分钟的短片大约需要 60-100 个镜头，每个镜头平均生成 3-5 次才能选出可用的。

第八步：提示词写作模板

[样式前缀]

[角色名]: [外形关键词 1-2 个], [动作描述], [情绪描述], [环境].

Camera: [从 Higgsfield 提示词库选取的摄影机运动].

[可选：锚点结果描述，代替过程描述]

常用摄影机词汇（直接从 Higgsfield Academy 提示词库复制）：

• 静止：The camera stays planted in one immovable position. Zero motion — no drift, no shake.
• 推进：One continuous decisive dolly in on straight ground rails — pure forward travel at constant speed.
• 上摇：Pure TILT UP: the camera rotates vertically upward at one constant smooth speed.
• 环绕：A smooth constant-speed circular drone flight around the subject, [8]-meter radius.
• 跟随：Handheld follow shot, medium distance, smooth stabilized motion tracking the subject.

第九步：难镜头策略

动作场景 → 用锚点式提示词（Zephyr 技巧 2）：

[样式前缀]

[角色名] executing a high-risk combat maneuver. 
Goal: convey overwhelming force, split-second precision, and barely-controlled desperation.
Camera: low angle, dynamic handheld follow.

不要描述每一个动作步骤。让模型自己填充。

物理异常场景 → 修改参考图输入（Zephyr 技巧 1）：

• 倒置 → 把参考图上下翻转再上传
• 水中 → 在参考图上叠加蓝色半透明滤镜再上传
• 受损 → 在参考图上直接涂改再上传

特写武器操作 → 上传独立特写参考图： 把武器局部放大到整张图的 70% 以上，再配提示词。

第十步：管理失败镜头

每个镜头生成 3-5 个版本是正常的。建立一个命名规则：

S[场次编号]_C[镜头编号]_v[版本号]_[KEEP/FAIL]
例：S02_C04_v3_KEEP

保留失败镜头，不要删。Zephyr 团队把 bloopers 也放进了开源资产，因为失败记录是最好的工程文档——你回头看的时候，能准确知道这条路走不通的原因。

---

阶段五：剪辑与收尾

第十一步：时间线组装

建议使用 DaVinci Resolve（免费版足够）。流程：

1. 把所有 _KEEP 镜头导入，粗剪出完整故事骨架
2. 用色彩校正统一跨镜头的色调（补充样式前缀没有覆盖到的差异）
3. 加音乐和音效（AI 生成音乐：Suno / Udio，音效库：Freesound）
4. 最终输出 4K，导出时保持原始帧率

第十二步：把你的项目文件开源

Zephyr 团队选择开放所有资产，包括失败镜头。这不只是慷慨，这是方法论：

• 你的开源项目文件就是你的技术简历
• 社区会基于你的资产做 remix，带动二次传播
• Higgsfield Film Festival 的评委也在看制作过程，不只是最终视频

在 Cinema Studio 里把项目设为公开，写一份和 Zephyr 一样的「Project Brief」，把所有技术决策和弯路都记录下来。

---

资源和时间估算

阶段	独立创作者	2-3 人小团队
前期开发	3-5 天	1-2 天
角色资产制作	3-7 天	1-3 天
正片生成	2-4 周	1-2 周
剪辑收尾	3-5 天	1-2 天
合计	4-7 周	2-4 周

生成次数估算（5 分钟短片）：

• 镜头数量：80-120 个
• 每镜头平均尝试次数：3-5 次
• 总生成次数：约 300-600 次

降低成本的技巧：

• 先用文字描述验证构图，确认没问题再生成
• 在 Cinema Studio 用最低参数快速出一版草稿，确认方向后再高质量出正式版
• 场景一致的镜头批量生成（同一参考图 + 同一样式前缀，换摄影机运动）

---

从 Zephyr 开源资产开始

最快的入门路径：

1. 打开 https://higgsfield.ai/original-series/zephyr-special/full-film
2. 点击「Open project」进入 Cinema Studio
3. 看 Higgsfield 团队每个镜头用了什么参考图、什么提示词
4. 找一个你感兴趣的镜头，把它的参考图和提示词改成你自己角色的，重新生成

你不需要从零开始——站在开源项目的肩膀上，就是他们开源的意义。

---

Mycelium Protocol — 追踪 AI 系统的底层演化

---

关于 Mycelium

菌丝协议。持续追踪 AI 工具、系统和实验的内容节点。

---

Higgsfield Zephyr: Special — Fully Open-Source K-Pop Mecha AI Film with Complete Production Breakdown

by Mycelium Protocol

---

This is the first time in the AI video space that someone has torn open an entire mecha K-pop short film and shown you exactly how it was made: every prompt, every character reference, every failed take alongside the final cuts — all public.

ZEPHYR: Special, produced by Higgsfield, runs 5 minutes 7 seconds, built on Seedance 2.5, and hit 11,000+ views within 16 hours of release.

Full project files: https://higgsfield.ai/original-series/zephyr-special/full-film
(Open in Cinema Studio to browse all prompts and assets)

---

The World of Zephyr

In the Zephyr universe, five girls are simultaneously K-pop idols and elite mecha pilots:

• MIRA — strategic leader, mech carries a flamethrower
• REINA — close-combat specialist, explosive power
• NAOMI — technical specialist, mech equipped with a grappling hook; central perspective of Special
• ZERO — calm sniper type
• HARUMIN — support role, the squad’s emotional anchor

Special episode: Naomi is in dire straits, her mech trapped upside-down. A mysterious guest character with mech combat mastery beyond the squad’s expectations rescues her.

---

Two “Impossible Shots” — and How They Did Them

The production team disclosed the engineering challenges and solutions behind two shots in the open-source brief.

Technique 1: The inverted character sheet

Naomi’s mech is upside-down in the opening action. The team needed to convey disorientation and shifted gravity.

First attempt: precise prompting to describe the physical state. It failed — the model didn’t accurately reproduce the upside-down physics.

Solution: invert the input, not the prompt. They flipped the character reference sheet so the physics were baked into the input stage. The prompt no longer needed to explain why the mech was inverted — it could focus on the emotional and visual details instead.

Result: Naomi now has a dedicated “inverted state” character sheet.

Engineering assumption: The model is better at inferring physical states from visual inputs than reconstructing them from text descriptions. Encoding physics into the image input bypasses the text-to-physics translation loss.

Technique 2: Anchor-point minimal prompting

Naomi’s rescue from the mech needed to convey danger, the immense weight of the mech, Naomi’s physical capability and fearlessness — without losing her grace and vulnerability.

The team opted for less strict prompting, giving the model more creative control by specifying only key anchor points:

We wanted to convey a sense of danger, the massive weight of the mech, and Naomi’s physicality and fearlessness — without losing her grace and vulnerability.

The model filled in the action details within those constraints and produced something the team couldn’t have generated with precise frame-by-frame description.

Engineering assumption: In complex action sequences, overly precise prompts compress the model’s creative space and produce stiffness. Specifying the desired outcome (emotion/physical result) rather than the process yields more natural motion.

---

What’s Open-Sourced

ZEPHYR: Special has published 197 assets, including:

• All prompts used (no exceptions)
• Final cuts that made the edit
• Failed takes and bloopers — including dead-end approaches
• Character reference sheets (including special states like the inverted sheet)

Anyone can open the project in Higgsfield Cinema Studio and see the complete generation history, from the first draft to the final version.

---

Episode 1 Engineering Lessons (Seedance 2.0)

The team also shared lessons from the original Zephyr episode (made on Seedance 2.0, when they were still establishing the pipeline).

The mech character sheet trap

Initial approach: a single “master sheet” showing all mech details — weapons, cockpit mechanism, all states.

Problem: if the model sees a detail, it will try to show that detail. The open hatch mechanism, visible on the character sheet as a state illustration, kept appearing even in scenes where the hatch should have been closed. The two states conflicted, deforming the overall mech design.

Solutions:

1. Separate character sheets per state: maintain distinct sheets for weapons retracted vs. extended. Only include the weapon on a sheet when it appears in the scene.
2. Close-ups as standalone inputs: for specific weapon operations (Mira’s flamethrower extending), upload a close-up of just that component as a separate input — don’t rely on the model zooming into a full-body sheet.
3. Text annotations on character sheets are nearly useless: labels like “retractable revolver in arm” don’t get interpreted as instructions. You still need to describe the full weapon action in the prompt.

Style prefix discipline

Episode 1 lacked a unified style prefix, leading to visual inconsistency across scenes. Recommendation: establish a consistent style prefix (lighting style, color temperature, camera language) before production begins, and open every prompt with it.

---

Character Creation Pipeline

Three stages:

Stage 1 — Character development: brainstorm with Higgsfield Supercomputer or Claude. The team explicitly says: “you can easily get that foundational layer of depth just by brainstorming with AI — don’t let a lack of knowledge stand in the way.” Knowing a character’s personality is also a generation guide: it makes it far easier to write their acting, reactions, and story arcs.

Stage 2 — Costume and appearance: Higgsfield Soul 2 for character and costume creation. Soul 2’s fashion capabilities handle any style combination, making it ideal for the idol × warrior aesthetic.

Stage 3 — Character sheet refinement: Nano Banana 2 Pro and Seedream to compile and refine designs into final reference sheet assets.

---

What Seedance 2.5 Changed

Zephyr Episode 1 was built on Seedance 2.0 (just released, heavy experimentation). Special runs on Seedance 2.5. The team describes the difference directly:

Today, AI creators are gaining an experience working with actors that closely mirrors being on an actual set. Technical workarounds and wrestling with models are taking a backseat.

The two “impossible shots” were achievable without compromise specifically because Seedance 2.5’s character consistency and scene control had advanced far enough to execute them.

---

The Platform Stack

Higgsfield is building an AI-native content production platform, not just a generation tool:

Tool	Role
Cinema Studio	Scene control, project management, full generation history
Soul 2	Character / costume design
Nano Banana 2 Pro	Image refinement
Seedream	Character sheet compilation
Seedance 2.5	Primary video generation model (ByteDance)
Supercomputer	AI-assisted brainstorming
Academy	Production pipeline courses (free)

Hell Grind (also open-sourced simultaneously) provides another reference point: 365,561 views, 115,446 generations, 3,138 public assets — a complete parallel production case study.

The Higgsfield Global Film Festival ($1M prize pool) is running now. The timing of these open-source releases is deliberate: share the production craft to help entrants understand what separates technically accomplished AI filmmaking from content generation.

---

Seven Engineering Principles

Extracted from the full Zephyr open-source project:

1. Encode physics in the input, not the prompt — inverted character sheet technique
2. Complex action scenes: anchor points, not process description — let the model fill in movement details
3. Multi-state character sheets — separate sheet for each important state (retracted/extended/special)
4. Close-up inputs for detail operations — don’t expect the model to zoom into a full-body sheet
5. Style prefix for visual coherence — establish consistent lighting/tone/camera language before production
6. Character first, then generation — personality drives performance, reactions, and arc
7. Publish the failures — bloopers reveal where the real engineering problems were, more clearly than successes do

---

---

Step-by-Step Tutorial: How an Individual or Small Team Makes a Sci-Fi Blockbuster

Prerequisite: A Higgsfield account (free to register) with Seedance 2.5 access. Open the Zephyr project files alongside this guide: https://higgsfield.ai/original-series/zephyr-special/full-film

---

Phase 1: Pre-Production — 2–3 Days

This is the highest-ROI phase. Skipping it doubles your cost in the production phase.

Step 1: Develop your world and characters with Claude or Higgsfield Supercomputer

Paste this template into Claude:

I want to make a 3–5 minute sci-fi animated short.

Setting: [one sentence — e.g. "2087 wasteland city where humans and mechanical life coexist"]
Number of protagonists: [1–3; solo creators should target 2 max]
Central conflict: [one sentence — e.g. "an amnesiac mercenary discovers they were engineered as a weapon"]

For each character, generate:
1. Appearance keywords (3–5, for image generation)
2. Personality traits (3–5, for writing performance descriptions)
3. A signature gesture or habit (1–2, for distinctive shot language)
4. The tension in their relationship with other characters

Why this first: the Zephyr team stated clearly — “knowing your characters makes it much easier to write their acting performance, reactions, and potential story arcs.” Your character doc is the reference for every prompt you write — not just backstory.

Step 2: Define your style prefix

Before generating a single frame, lock in your visual language. Write the following as a fixed block that opens every prompt you write:

[Example style prefix]
Cinematic, 4K, anamorphic lens, neon-noir lighting, cool blue and amber color grade,
shallow depth of field, film grain, [your world name]

The original Zephyr episode had no unified style prefix, causing visual inconsistency across scenes — an explicit mistake the team called out. Your style prefix is the film’s visual bible.

---

Phase 2: Character Asset Production — 2–5 Days

Step 3: Generate character appearances with Higgsfield Soul 2

Soul 2 is built for character and costume design. Open Higgsfield → Image → Soul 2.

Two generation rounds per character:

Round 1: Find the look

[gender] [hair color] [hairstyle], [3 appearance features], [costume style],
[accessories or signature element], sci-fi [your style words], full body,
white background, character design sheet

Round 2: Lock the hero image
Pick the best result and iterate in Soul 2 until you feel “this is them.” This image is the anchor for everything that follows — do not change it arbitrarily.

Step 4: Build multi-state character reference sheets

This is where the Zephyr team accumulated the most painful lessons. Each character needs at minimum:

Sheet type	Use	Note
Standard standing (front + side)	All regular scenes	Primary reference
Emotion close-up (2–3 expressions)	Dialogue / emotional scenes	Upper body only
Action pose	Combat / running scenes	Do NOT include weapons
Weapon / prop extended (one sheet each)	Weapon-use scenes	Upload only when needed

Use Nano Banana 2 Pro to refine each sheet — sharper details, more stable output. Use Seedream to combine multiple design drafts into a unified-style reference set.

Critical rule: weapons and mechanisms must never appear on the standard standing sheet. If the model sees a detail, it will render it.

Step 5: Vehicle / mech design (if applicable)

Mechs require more state management than human characters:

Required separate sheets:

1. Full body standard form (all special parts stowed)
2. Cockpit area close-up — hatch closed
3. Cockpit area close-up — hatch open
4. Weapons system close-up (one sheet per weapon)
5. Inverted / damaged state (if the story calls for it) — use Zephyr’s inverted sheet technique

Rule: every component that opens, extends, or transforms needs two independent sheets: stowed state and deployed state.

---

Phase 3: Pre-Production Asset Testing — Half a Day

Test your reference sheets before committing to full production. Catch problems here, not mid-shoot.

Step 6: Stability testing in Cinema Studio

For each character, generate 3–5 test shots using this template:

[Style prefix]

[Character name], [2–3 appearance keywords], standing in [neutral environment],
looking at camera, medium shot, cinematic.

Camera: static shot, locked tripod.

Check:

• Is the character’s appearance consistent across shots (hair color, costume, face)?
• Did any unexpected details appear (weapons, mechanisms, extra elements)?
• Does the lighting and color match your style prefix?

If inconsistent → back to Step 4 to adjust the reference sheet, then re-test. Only proceed to production after passing this check.

---

Phase 4: Principal Photography — Shot by Shot

Step 7: Break the script into a shot list

Don’t think in scenes — think in shots. Each shot list entry:

Shot [number]
- Character: [who]
- Action: [what they do]
- Emotion: [what it feels like]
- Camera: [position + movement]
- Duration: [3–8 seconds]
- Reference sheet: [which one]
- Special technique: [inverted / anchor-point / close-up input / etc.]

A 5-minute short film requires roughly 80–120 shots. Budget 3–5 generation attempts per shot to find a keeper.

Step 8: Prompt template

[Style prefix]

[Character name]: [1–2 appearance keywords], [action description], [emotional quality], [environment].

Camera: [camera movement from Higgsfield Academy prompt bank].

[Optional: anchor-point outcome description instead of process description]

Camera vocabulary — copy directly from the Higgsfield Academy prompt bank:

• Static: The camera stays planted in one immovable position. Zero motion — no drift, no shake, no breathing.
• Dolly in: One continuous decisive dolly in on straight ground rails — pure forward travel at constant speed along the axis.
• Tilt up: Pure TILT UP: the camera rotates vertically upward at one constant smooth speed.
• Orbit: A smooth constant-speed circular drone flight around the subject, [8]-meter radius.
• Follow: Handheld follow shot, medium distance, smooth stabilized motion tracking the subject.

Step 9: Hard shot strategies

Action sequences → anchor-point prompting (Zephyr Technique 2):

[Style prefix]

[Character name] executing a high-risk combat maneuver.
Goal: convey overwhelming force, split-second precision, and barely-controlled desperation.
Camera: low angle, dynamic handheld follow.

Don’t describe every step. Let the model fill in the movement.

Physically abnormal scenarios → modify the reference image input (Zephyr Technique 1):

• Upside-down → flip the reference sheet vertically, then upload
• Underwater → add a blue semi-transparent overlay to the reference sheet
• Damaged / destroyed → paint over the relevant parts of the reference sheet

Weapon operation close-ups → upload a dedicated close-up input: Crop the weapon to fill at least 70% of the image before uploading.

Step 10: Managing failed takes

Generating 3–5 versions per shot is normal. Use a consistent naming convention:

S[scene]_C[shot]_v[version]_[KEEP/FAIL]
Example: S02_C04_v3_KEEP

Never delete failed takes. The Zephyr team included bloopers in their open-source assets for this reason: failure logs are the best engineering documentation. When you look back, you know exactly why that approach doesn’t work.

---

Phase 5: Edit and Delivery

Step 11: Assemble the timeline

Recommended tool: DaVinci Resolve (free version is sufficient). Steps:

1. Import all _KEEP shots, cut a rough assembly of the complete story
2. Color grade to unify cross-shot tone variations (fill in the gaps your style prefix didn’t cover)
3. Add music and sound design (AI music: Suno or Udio; SFX library: Freesound)
4. Export at 4K, preserving native frame rate

Step 12: Open-source your project files

The Zephyr team published everything — including bloopers. This is methodology, not just generosity:

• Your open-source project file is your technical portfolio
• The community will remix your assets, driving organic reach
• Higgsfield Film Festival judges review the production process, not just the final cut

In Cinema Studio, set your project to public and write a Project Brief the way Zephyr did: record every technical decision and every dead end you encountered.

---

Resource and Timeline Estimates

Phase	Solo creator	2–3 person team
Pre-production	3–5 days	1–2 days
Character assets	3–7 days	1–3 days
Principal photography	2–4 weeks	1–2 weeks
Edit and delivery	3–5 days	1–2 days
Total	4–7 weeks	2–4 weeks

Generation volume estimate (5-minute short):

• Shots required: 80–120
• Average attempts per shot: 3–5
• Total generations: approximately 300–600

Cost reduction tactics:

• Verify composition in text first before generating
• Use lowest settings for a quick draft, then generate the final version at full quality only after confirming direction
• Batch-generate shots in the same environment (same reference sheet + same style prefix, varying only camera movement)

---

The Fastest Starting Point: Use the Zephyr Assets Directly

1. Open https://higgsfield.ai/original-series/zephyr-special/full-film
2. Click Open project to enter Cinema Studio
3. Browse the Higgsfield team’s reference sheets and prompts, shot by shot
4. Find a shot you want to adapt — swap the reference image for your character, modify the prompt, regenerate

You don’t need to start from zero. Standing on the open-source project is exactly the point of open-sourcing it.

---

Mycelium Protocol — tracking the deep evolution of AI systems

© 2026 Mycelium Protocol. All rights reserved.