飞升 · 视觉音乐概念专辑Ascension · A visual music concept album

一张与 AI 共同创作的视觉音乐概念专辑A visual music concept album, made with AI

Ascension

一张 4060 Ti，0 成本^*，4 天生成。
让想象力飞升。One 4060 Ti. Zero cost^*. Four days.
Let imagination rise.

概念与分镜 1 天，素材生成 2 天，自动化剪辑 1 天。
我只写提示词。网站与制作方法，免费开源。One day to design. Two to generate. One to edit.
I wrote the prompts. The site and process are free to share.

免费拿走制作方法Get the free making-of kit

章浩 / ZHANG HAO

需要什么配置？What do you need?

这次用一台普通电脑完成。通过量化模型、动态显存、NVMe 卸载和逐镜头生成，让大模型在有限显存里跑起来。Made on a consumer PC. Quantized models, dynamic VRAM, NVMe offloading and one shot at a time made the larger models manageable.

4060 Ti / 8GB显卡与显存GPU and VRAM

32GB系统内存System RAM

NVMe SSD模型存储与卸载Model storage and offloading

这是本项目已跑通的配置；尚未验证更低配置。生成速度随分辨率、时长和步数变化。This configuration worked for this project. Lower specifications have not been tested. Speed varies with resolution, duration and sampling steps.

模型、参数与实测耗时Models, settings and measured times

留存环境：ComfyUI 0.37.4 · Python 3.12.9 · PyTorch 2.9.1 / CUDA 13.0。逐镜头串行生成，batch = 1。Recorded environment: ComfyUI 0.37.4 · Python 3.12.9 · PyTorch 2.9.1 / CUDA 13.0. Sequential generation, batch = 1.

--enable-dynamic-vram --fast-disk
--disable-pinned-memory --vram-headroom 2.5
--cache-ram 6 8 --preview-method none

用途Purpose	模型Model
I2V / R2V	MiniMax H3 FL2VA / REF2VA · INT8
文本与参考编码Conditioning	Qwen3-VL 32B · NVFP4 / AWQ
解码Decoding	H3 Video VAE INT8 · Audio VAE FP32
快速预览Fast previews	H3 Turbo 8-step LoRA
风格实验Style studies	Realism People · Looping Sketch Anime

v11 风格试片：640×384，73 帧，24fps，8 步；单段实测 2.50–3.58 分钟，中位数 2.69 分钟。v10 高清镜头：1280×736、25 步，输出裁切为 720p。两组使用不同配置，不能套用同一耗时。V11 style studies: 640×384, 73 frames, 24fps, 8 steps; measured 2.50–3.58 minutes per clip, median 2.69. V10 HD footage used 1280×736 at 25 steps, cropped to 720p. These presets do not share the same generation time.

官方安装指南 Official setup · 9 项模型与校验记录 Nine model records & hashes

先让故事成立。Give it a story.

她是红衣信使。为了让静止的城市重新流动，她带走光种，遭遇追逐，最后主动跃入深渊。不是被动获救，而是用一次选择改变世界。A courier in red takes a seed of light to awaken a frozen city. Pursued and cornered, she chooses to leap. Her decision changes the world.

短发、红衣、金色腕线固定角色；“英雄之旅”把出发、试炼、转变和归返串成因果。A short bob, red coat and gold wrist thread anchor her identity. A hero’s journey connects departure, trials, transformation and return.

看见异常The call夺取光种The threshold追逐与抉择The ordeal跃下与飞升Transformation携光归返The return

把想法变成分镜。Turn ideas into frames.

把角色、动作、景别、光线和前后关系写进提示词。借助技能约束构图与连续性，再通过文生图逐轮确定画面。红绸、人物、调色和风格实验，留存的设计稿都可以在这里浏览。Prompts specify character, action, framing, light and continuity. Skills guide composition; image generation develops the look. Browse the preserved designs, from the ribbon concept to characters, grading and style studies.

先有故事 / A story first — 先有故事A story first

再有选择 / Then a choice — 再有选择Then a choice

最后是飞升 / Then, ascension — 最后是飞升Then, ascension

查看提示词结构See the prompt structure

同一角色与服装一个明确动作景别和机位材质与光线道具、视线与运动方向避免的错误。
例如：红衣信使在断桥边握紧光种，三分之四侧面中近景，干燥哑光材质；她主动看向深井，人物保持站立，不能画成失足坠落。Consistent identity one clear action framing and angle materials and light props, eyeline and screen direction errors to avoid.
Example: the courier grips the seed at the broken bridge, in a three-quarter medium close shot with dry matte materials. She deliberately looks into the shaft, standing upright rather than slipping.

用到的技能与安装方式Skills and installation

概念与分镜Concept & storyboards	agentic-video-foundry · remotion-video · remotion-best-practices
节拍与构图Rhythm & composition	beat-sync-editing · shot-composition
剪映界面验证Editor verification	cua-driver
网站共创Website design	grill-me · brainstorming · i-have-adhd · writing-plans

技能负责给 AI 工作方法；模型负责生成画面。仓库附带本项目的可安装技能和第三方来源说明，网站技能与 MV 制作技能分开记录。Skills guide the agent; models generate the visuals. The repository includes installable project skills and third-party provenance, separating website work from MV production.

技能清单与安装 Skill list & setup

批量生成，逐条筛选。Generate. Review. Keep.

给每个镜头写动作与运镜提示词，把参考分镜送进 ComfyUI。模型按队列逐段生成，AI 检查主体、动作、连续性和画面故障，区分可用片段、需重做片段与废片。Write motion and camera prompts for each shot, then feed the reference frames into ComfyUI. The queue generates clips one by one. AI reviews identity, motion, continuity and visual defects to select, retry or reject them.

分镜与参考帧Storyboard and references

动作与运镜提示词Action and camera prompt

ComfyUI / H3

抽帧验证与取舍Frame review and selection

公开 68 份节点工作流与 68 份 API 配置，75 张原始输入图可单独下载。角色漂移与动作错误的修正记录，也一并公开。The archive shares 68 node workflows and 68 API configurations, with 75 original input images available separately, alongside corrections for identity drift and motion errors.

让剪辑跟着音乐呼吸。Let the edit breathe
with the music.

我提出要求，AI 执行制作。它通过剪映 MCP 创建工程、导入素材和调整时间线。分析音乐起音与强弱变化，再检查每个切点前后的画面：动作有没有接上，情绪有没有被打断，高潮有没有真正落下去。I gave the direction; AI executed the production. Jianying MCP created the project, imported footage and shaped the timeline. Music onsets guide the cuts; frames on either side reveal whether actions connect, emotion carries and the climax lands.

长镜头铺垫，短镜头推进。无声节奏示意。Long shots hold. Short cuts build. A silent rhythm study.

实际切点抽帧对照。制作综合使用 MCP、脚本、接口与界面操作；并非每一步都通过鼠标点击。An actual cut-point review sheet. Production combined MCP, scripts, APIs and UI control, rather than using mouse actions for every step.

这支片子，
你也可以接着做。Take this film
somewhere new.

我把一路试过的想法、画过的分镜和跑通的方法，整理成了一本公开的制作笔记。你可以翻着看，也可以挑一个镜头，换成自己的故事。The ideas, sketches and settings that made this film are here to explore. Browse the notebook, or choose a shot and give it your own story.

去 GitHub 拿走这份项目 Get the project on GitHub

Four days: one to design, two to generate, one to edit

我来提想法，
AI 把它做出来。I set the direction.
AI did the production.

先聊清楚主角想做什么，再画分镜。让静态画面动起来之后，逐条挑选，跟着音乐剪，回头看每个切点有没有接上。First, decide what the character wants and draw the scenes. Generate the motion, choose the usable takes, edit to the music, then check each cut.

不是一句话就完成。我反复提要求、选方向，AI 反复生成、修改和检查。It took more than one prompt. I kept giving direction and choosing versions; AI kept generating, revising and checking.

翻翻完整制作记录 Read the production notes