本文来源:https://mp.weixin.qq.com/s/xfx8RV5bs0fqSa0ZwcdfnQ
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
📦 3 Parts + Conclusion
👉 滑动
PART 01
OpenCode
coding agent
PART 02
NVIDIA NIM
免费的LLM API
PART 03
Agnes
免费 AIGC 三件套
PART ///
写在最后
下一步
token 是AI最大的成本
失业之后,我一直在找副业的方向。从写自动化脚本、做工具,到后来vibe coding做机型,再到后来玩本地AIGC。折腾了大半年,才弄懂免费才是生产力。
生产力来源于赚钱的目标,副业/创业 最大的难题来自于成本,本来已经没有收入了,如果卡不住成本,就没法充分试错,就没法成功。
能不能把 token 成本降下来,几乎决定了副业能不能跑出来。这篇文章我将介绍我的一些免费实践:一个开源 coding agent + 一个免费LLM API + 一个免费 AIGC 三件套,小成本也能把 vibe coding 跑出MVP。
御三家的大模型虽然好,不过架不住成本高啊,在没有明确的赚钱前提下,我觉得还是成本更加关键,现在的免费模型也比较强了,很多时候不是卡在代码产出速度不够快,而是打通客户和流量的瓶颈。
轻量创业的本钱就是时间和算力。算力端如果压不住,赚的还没烧的多,最后只是被收割的对象。所以比起挑"最强"的模型,先把"够用且免费"的模型拼出来,是我现在的第一优先。
01
PART
OpenCode:终端里的免费 coding agent
TERMINAL · FREE CODING AGENT
它是什么
OpenCode 一个开源的 AI coding agent,你可能用过claude code,codex,cursor这些顶级agent,OpenCode 相比之下略逊一筹,不过重在它开源,且接入各方api格外方便,最大的好处在于可以一直用免费模型。虽说是coding agent不过能做的事情很多,例如我这篇公众号的排版也是用OpenCode跑摸鱼小李的skill开展排版的。
怎么装到本机
官网:https://opencode.ai/
它有CLI端,vscode插件,IDE端,我比较习惯的是使用IDE端。直接下载之后,想做的事情和它对话完成就行,我觉得AI年代了不用写什么具体教程了,都教给agent就好。
它最关键的一张牌:自带免费模型

OpenCode经常会更新免费使用大模型,常驻的例如Big Pickle,Nemotron3可能有些不太好用,最近一直免费了有超过一个月的deepseek v4 flash和MiMo V2.5是我最常用的模型,我用这两个模型有写过至少5个项目了。之前在做一个公司爬虫项目的时候,也是一直用的MiMo,实测还是可以比较从容的完成任务的。
当然这些模型是阉割过的,一般上下文会砍到200K左右,但合理使用还是够用的。
✦ 我的用法日常最常用的几个情形:写脚本、改 bug、起新项目(web,app,pc应用)、写爬虫、逆向分析等;opencode很吃 AGENTS.md 的设置,先设定好开发测试流程,可以事半功倍。
为什么不用TRAE,WorkBuddy(了解更多),Qoder这类国产agent?
最大的难题是免费额度有限,我去年一直用TRAE,免费模型比较多,不过TRAE用起来改bug经常是改来改去没有结果,几个免费国产模型来回切换也很费劲才能应对难题,一次偶然机会试了opencode之后,不过配合mimo应对难题格外顺畅,后来我就把TRAE卸载了,再后来听说TRAE的免费额度政策改了,给的已经很少了。
wb,我之前用这个开发一个app应用,太重了,用起来不灵活,而且后来因为代码不在本地无法打包的难题中断了,我后来搞清楚写代码还是要用cb才行,机型线的定位和宣传我觉得也有点难题。
在其他的实际上都差不多了,也是会给一些免费额度,不够用的,切来切去很麻烦,没必要了,Qwen一开始也是很强的,现在慢慢掉队了,剩一些开源模型还是很强的。
用好几个md和skill
开发项当下先找好需要用到的skill,设置好agent.md是重中之重,提效格外多,然后就是要用design.md之类一系列md文件来约束。
opencode我之前用过omo插件,实际用下来不太理想,就没去深究了,因为oc更新太快了,有时候一天几版,所以我基本还是靠原生能力。
至于我做出过哪些小机型/demo,下次有机会再介绍,AI工具更新太快,很多时候就是做着做着就过时了,所以后来我索性就让自己慢下来,沉淀一些复古的东西。
02
PART
NVIDIA NIM:老黄真是个好人啊!
FREE INFERENCE · OPEN MODELS
NIM 是什么
NVIDIA 把一批开源大模型封装成 OpenAI 兼容的 API,供开发者免费调用,托管在 build.nvidia.com 上。
模型很多,不限制token量,不过限制token速度。
PS:用冷门一点的旧一点的模型会快很多
怎么注册
进 build.nvidia.com,用 NVIDIA 账号登录,然后到 "API Key" 页面自己 Generate 一个。把这个 Key 记下来——下一步要用。
接入方式--直连
像opencode这么先进的agent,直接填入key就可以用了。
我用的比较多的GLM-5.2和deepseek,minimax实际上我用2.7版本比较多,众所周知,M3之后的版本变得太慢了太墨迹。2.5/2.7以前的版本真的是傻快傻快的,除了不那么聪明,应付小工作格外方便,例如接去养马养虾,不过我一直没找到养马养虾的使用情形。kimi2.6最近被下掉了,估计在准备上k3了。
第二种接入——CC-Switch
CC-Switch 是一个跨 Claude Code、Codex、OpenCode、Claude Desktop、Gemini CLI、Grok Build、OpenClaw、Hermes 这一票编程 CLI 工具的"规格集中管理器"。它本身是个 Tauri 桌面 app,官方站是 ccswitch.io。
claude code和codex,就需要使用cc-switch插件了,我之前还用到另一个cli-proxy-api的插件做格式转换,cc-switch升级之后可能自带了,我不确定。有段时间没用了。
codex的话,目标达成性强一些,经常一个任务跑一个小时。cc比较发散,如果提示词或者一些harness没做好,可能会给你改的面目全非,我以前有过几次回滚代码,要用好需要一些学习成本。
我不太迷信哪个工具更强,更多的时候就是看自己跑几个需要跑下来的感觉,最近的TRAE我也用了几次,比之前好很多了,不过因为限额的难题,我没作为主力使用。
✦ 我自己的用法在 CC-Switch 里给 Claude Code和codex 配 NIM 的 GLM / DeepSeek / MiniMax,两边都能直接 vibe coding,零 token 成本,而且上下文不缩水。
03
PART
Agnes:免费的 AIGC 三件套
TEXT · IMAGE · VIDEO
生图、生视频、文本
Agnes 是一个新加坡公司的大模型机型,主打文本、图像、视频三方向的生成能力,三个方向都给出一个免费的flash版本。
直接给出了一个网页版本叫pavo,网址:https://app.pavo-ai.work/
到底怎么样
很多博主都吹这个免费模型怎么好用,不知道是不是都恰饭的,但毕竟免费给出,咱不能说它不好,具体怎样,我做了一些真实的测试。
本来我把这个免费api直接集成到skill里面,让ai依据上下文推提示词给它直接生图的,后来成效实在不太能用,就下掉了,真要批量搞还是得上gptimage2。
生图
生成一个女仆装的少女正在跳热舞的图片
生成一个古风美女,抱着一个琵琶,背景是雪地
带有绿树、浅色围墙、街道设施的绿意街道,充满阳光明媚春日感,暖调阳光洒下形成光晕,头发丝和脸上有光影。人物为18岁漂亮模特,长发自然垂落肩膀,一只手做出比耶的动作,另一只手自然垂落,人物脸部不变形,画面清新自然、高清,氛围开心,细节极致。比例 9:16
少女感网拍,冷白皮超美女孩,清透裸妆,户外情形,仰拍镜头,正脸看镜头,蓝天白云,纯欲感,白嫩嫩的皮肤,如琼玉般润滑,天蓝色色慵懒针织外套,氛围感,姿态慵懒,总体姿态传递出一种沉浸在自我风格中的独特韵味,尽显软萌气质,高颜值,19岁女生尖下巴,三庭五眼比例均衡,身材娇小,娇羞的表情,细腻陶瓷水嫩肌肤,偏冷白皮,随机的妆容,随机发型和颜色,随机的饰品,随机的动作,侧光,发丝轻飘,凌乱,随机角度拍摄,随机pose,总体有很多的不确定性,但又好看
制作一张高分辨率的方形 (1:1) 食物海报,内容是 (您的餐点)。以动态的、爆炸式的构图捕捉这道菜,使其仿佛向观者迸发。展现逼真、令人垂涎的质感和鲜艳的色彩。将其置于深黑色背景前,辅以专业的戏剧性摄影棚灯光、多元的景深和高比较度。强调其高端、奢华和视觉冲击力。排除任何手、人物、文字、标志或分散注意力的餐具。
风格:写实 CG 科幻恐怖电影截图,《杀戮都市:0》暗黑末世风格,高度还原原作怪物质感与战场氛围媒介:电影级 CG 渲染,4K 超高清分辨率,细节锐利,光影冷峻,无噪点主体怪物:滑瓢鬼尸魔终极形态外观与材质:通体为暗红肌肉与外露银白骨骼,皮肤呈枯槁撕裂状,肌肉纹理与骨骼轮廓清晰可见,表面带有湿润的血渍与磨损痕迹肢体特征:背部生有多根巨大骨骼尖刺,骨刺呈灰白骨质感,边缘锋利且布满裂纹;手肘处延伸出弯曲骨刃,形态狰狞;头部为枯槁恶魔面容,眼窝深邃,面部骨骼突出,表情冷漠且充满压迫感动态姿态:身体微微前倾,右臂向前伸出,指尖精准点向人类角色额头,姿态傲慢而充满威慑力;左臂自然垂落,提着一颗人头,手指呈枯爪状,关节扭曲,总体姿态彰显绝对力量长处人类角色:东亚男性战士发型与面部:黑色利落短发,发丝紧贴头皮,面部带有细微伤痕与血迹,眼神坚毅,抬头直视怪物,表情紧绷,展现出对抗的决绝服装与配饰:身着黑色紧身战斗服,材质为哑光橡胶与金属拼接,表面有细微磨损与划痕,颈部与肩部带有银色金属卡扣与管线装饰,凸显科幻战斗质感,面部有战斗留下的污渍动态姿态:身体微微后仰,姿态紧绷,双手自然垂于身侧,肌肉线条隐约可见,展现出面对巨型怪物时的警惕与不屈环境与情形背景:夜晚大阪都市废墟,高楼大厦破损燃烧,浓烟弥漫,远处可见霓虹招牌与日文标识,灯光在烟雾中朦胧闪烁中景:破损的金属护栏横贯画面,地面布满碎石与废墟,远处街道可见零星火光与废弃车辆氛围:末世战场的压抑与绝望感,巨型怪物与渺小人类形成强烈视觉比较,凸显力量悬殊灯光与工艺光源:环境光为橙黄色火光与冷色霓虹灯光,主光源从侧上方打在怪物与人类身上,形成高比较度硬阴影,突出肌肉与骨骼的立体感质感:肌肉的湿润感、骨骼的粗糙感、战斗服的橡胶质感、烟雾的半透明感均高度还原,材质区分清晰
在阿尔卑斯山环境中,生成一张具有电影级奢华美学的高级时装冬季时尚肖像大片。主体为一位年轻女性,她坐在一块经过雕刻的冰块上,双腿自然分开,一只戴着冬季手套的手轻贴脸颊,神情放松又充满力量感。穿搭为:外层是黑色超大号羽绒服,下身搭配白色阔腿滑雪裤,脚穿配有红色鞋带的白色奢华冬季冬靴,头戴毛绒白色毛皮帽(遮住耳朵),手上佩戴冬季手套。情形设定在广阔的雪山景观中,前景是平滑起伏的雪堆,背景为壮丽巍峨的雪山,人物坐于雕刻过的冰块表面,天空是阴沉的冬日天空。光线使用阴天自然光,总体柔和、漫射,光源来自阴沉的天空。拍摄使用 85mm 镜头,浅景深,主体对焦锐利,为时尚编辑摄影风格。画面构图极简,人物居中,展现电影感画幅,总体是超洁净的奢华时尚大片风格。色调以冷蓝色、白色、柔和灰色与深黑色为主,氛围充满电影感、高端质感、宁静且富有力量。画面追求极致真实的质感,逼真展现服装面料细节与高度写实的冰块纹理,超高分辨率,照片级写实成效,对焦锐利,纹理细节清晰。
画面风格:数字插画与唯美风景绘画相结合的艺术风格,使用细腻柔和的笔触与多元的色彩层次渲染,总体展现浪漫梦幻的电影级光影质感,氛围宁静而充满诗意。核心元素:壮阔的海岸日落景观,前景盛开的红粉色玫瑰丛,中景连接岛屿的狭长岩石陆桥,远景矗立于葱郁山丘上的白色灯塔与深色针叶林,广阔平静的海洋与绚烂晚霞交织的天空。具体内容:画面底部被茂密的玫瑰花丛占据,红色与粉红色的玫瑰花瓣层层叠叠,纹理清晰可见,深绿色叶片穿插其间,茎秆挺拔自然。一条布满低矮花卉与青草的岩石陆桥从左下角蜿蜒延伸至画面中央的圆形小岛,岛上植被浓密,点缀着星星点点的各色野花,顶端矗立着一座简约的白色灯塔,塔身笔直,周围环绕着几棵高耸的深色柏树。远处海面波光粼粼,夕阳的余晖在水面铺就一条明亮的金色光路,海浪轻柔拍打礁石形成细腻的白色泡沫。天空展现由橙红向淡紫渐变的暖色调,云层稀薄柔和,光线温暖而弥散。构图方式:使用垂直画幅与纵深透视构图,前景花丛作为视觉框架引导视线向中景陆桥延伸,陆桥形成天然的对角线引导线将目光引向远景灯塔,背景海平线位于画面上方三分之一处,符合三分法原则。总体空间层次分明,前景锐利细腻,中景过渡自然,远景略带空气感虚化以增强纵深感。渲染参数设定为高精度数字绘图,使用宽色域色彩映射,表面反射率适中,岩石展现粗糙颗粒质感,水面使用平滑渐变与高光反射叠加,光影方向来自右上方低角度照射,形成柔和的侧逆光轮廓与温暖的体积光效,总体色调以暖橙色与青蓝色互补为主,营造出静谧唯美的黄昏海景意境。
画面风格:硬科幻电影级CG渲染风格,超写实建筑概念外观,色彩冷峻克制,以银灰、纯白和淡蓝为主色调,总体氛围宏大肃穆,具有强烈的将来工业美学与末世荒凉感,画质极致清晰,细节多元。核心元素:一座巨型将来主义机械设施,依附并悬挑于高耸的白色悬崖边缘,背景为广袤的荒原与远处的将来城市天际线。建筑主体由多层银灰色金属平台、深色合金管道、垂直支撑结构和精密机械组件构成,展现出厚重的工业质感与高科技感。悬崖表面洁白如雪或白垩岩,具有显著的垂直侵蚀纹理和层状结构。远处城市高楼林立,轮廓在大气透视中略显朦胧,天空展现苍白的高调蓝白色。具体内容:前景左侧是庞大的悬空建筑群,结构复杂,包含圆柱形管道、矩形舱室和网格状金属护栏。建筑边缘整齐切割,与天然悬崖形成人工与自然的强烈比较。部分重型机械装置悬挂在悬崖外侧,深入岩壁,表明出工程结构的复杂性。悬崖下方深渊不可见底,增强了高度感。背景中的城市建筑群密集排列,具备尖顶和流线型塔楼,暗示这是一个高度发达的文明遗址或前沿殖民地。光线均匀洒落,金属表面反射出柔和的高光,岩石阴影处展现冷灰色调。构图方式:使用竖构图,高角度俯视视角。主体建筑位于画面左侧三分之一处,视线顺着悬崖边缘向右后方延伸,引导至远景的城市中心。前景的机械结构与中景的垂直崖壁形成紧密的空间层次,远景的城市天际线给出环境背景,增强画面的宏大尺度感。工艺参数设定为全画幅传感器,使用24mm广角镜头,光圈f/8,快门速度1/500秒,ISO 100,白平衡日光模式,对焦模式为超焦距,确保从前景机械细节到远景城市轮廓均保持极致锐利,景深充足,画面无畸变,光影过渡自然平滑。
总结:比较吃提示词,提示词少的情况下大概率浮现多手脚,手指等缺陷,相同提示词的情况下表现不太赶得上开源的本地模型,和闭源差距很大。
生视频
图生视频:转为视频,自由发挥
用的刚才生成的图片
图生视频:转为视频,自由发挥,图中的女孩潇洒的饮酒作诗。
这里使用一张官方首页的一张示例图
图生视频:转为视频,自由发挥
模型的理解不太好,于是把提示词引导去掉,让模型自己来,确实好一点了。
总结:应该打不过本地的LTX2.3吧,而且没法微调,提示词理解维度有待提高。
文本
就不测了,都可以用GLM5.2或者deepseek了,这个没必要拉进来比了。
///
LAST
下一步:跑实际业务
NEXT
用好上面的白嫖三件套,很多业务情形已经可以打了,后面我应该会分享一些实际业务情形的应用中的一些经历。
先白嫖,再付费;先跑通,再放大
Source: https://mp.weixin.qq.com/s/xfx8RV5bs0fqSa0ZwcdfnQ If there are copyright issues, please contact us: 13951431913@139.com or leave a comment on the right.
Looking for a side hustle that requires buying a model membership upfront?
Free is productivity — the three-piece freeloading kit
Driving token cost to zero
OpenCode · NVIDIA NIM · Agnes — a zero-cost vibe coding loop
The token economics of an unemployed side-hustler
Free · hands-on
📦 3 Parts + Conclusion
👉 Swipe
PART 01
OpenCode
coding agent
PART 02
NVIDIA NIM
free LLM API
PART 03
Agnes
free AIGC three-piece kit
PART ///
Wrapping up
Next step
Tokens are the biggest cost in AI
After losing my job I have been looking for a side hustle. From writing automation scripts and building tools, to later doing vibe coding on products, to later playing with local AIGC. After half a year of tinkering I finally understood: free is productivity.
Productivity comes from the goal of making money, and the biggest difficulty in a side hustle or startup is cost. If you already have no income and cannot keep costs under control, you cannot afford enough trial and error, and you cannot succeed.
Whether you can bring token costs down almost determines whether a side hustle can work. In this article I will introduce some of my free practices: one open-source coding agent, one free LLM API and one free AIGC three-piece kit — enough to run vibe coding to an MVP at low cost.
The big three models are good, but they cannot stand up to the cost. Without a clear path to revenue, I think cost matters more. Free models are now quite strong too; often the bottleneck is not the speed of code output, but connecting to customers and traffic.
The capital of a lightweight startup is time and compute. If you cannot hold compute costs down, you earn less than you burn, and in the end you are just there to be harvested. So rather than picking the 'strongest' model, my first priority now is to piece together a model that is 'good enough and free'.
01
PART
OpenCode: a free coding agent in your terminal
TERMINAL · FREE CODING AGENT
What it is
OpenCode is an open-source AI coding agent. You may have used top-tier agents like claude code, codex or cursor; OpenCode is somewhat behind them, but the point is that it is open source and it is unusually convenient to plug in APIs from various providers — and the biggest benefit is that you can keep using free models. Although it is a coding agent, it can do a lot: the layout of this very WeChat article was done with OpenCode running my 'Slacking Li' skill.
How to install it locally
Official site: https://opencode.ai/
It has a CLI, a VS Code plugin and an IDE version. I personally prefer the IDE version. Just download it and talk to it about what you want done — in the AI era I do not think detailed tutorials are needed, leave it to the agent.
Its single most important card: built-in free models
OpenCode frequently updates its free model lineup. Permanent residents include Big Pickle and Nemotron3, which may not be that usable; more recently the free DeepSeek v4 Flash and MiMo V2.5, available for over a month, are the models I use most, and I have written at least 5 projects with them. When I was doing a web scraper project for a company, I also used MiMo throughout, and in testing it handled the task reasonably comfortably.
That said, these models are cut down — context is usually trimmed to around 200K — but used sensibly they are still sufficient.
✦ My usage — the most common everyday situations: writing scripts, fixing bugs, starting new projects (web, app, PC applications), writing scrapers, reverse engineering and so on. opencode leans heavily on the AGENTS.md setting, so defining your development and testing workflow up front roughly doubles your efficiency.
Why not TRAE, WorkBuddy (learn more), Qoder and other domestic agents?
The biggest problem is that the free quota is limited. I used TRAE all last year and it had quite a few free models, but with TRAE, fixing bugs often went round in circles with no result, and switching between several free domestic models was a hassle just to tackle a hard problem. After trying opencode by chance, pairing it with mimo handled hard problems unusually smoothly, so I uninstalled TRAE. Later I heard TRAE's free quota policy changed and now gives very little.
WorkBuddy — I used it to develop an app once, but it was too heavy and not flexible, and later I gave up because of the problem that the code was not local and so could not be packaged. I later worked out that for writing code you still need cb; as for the product line's positioning and marketing, I think that is a bit of a problem too.
The others are basically all the same — they give some free quota, and when it is not enough, switching back and forth is a pain, so there is no point. Qwen was strong at first but is falling behind; some of the remaining open-source models are still strong.
Using several md files and skills
For a development project, first find the skills you need — setting up agent.md is the top priority and helps enormously. Then use a set of md files like design.md to constrain things.
I used the omo plugin with opencode before, but in practice it was not ideal so I did not dig deeper, because oc updates too fast — sometimes several versions a day — so I basically rely on native capability.
As for what small products and demos I have built, I will write about them next time we have a chance. AI tools update so fast that half the time you are midway through building something and it is already outdated, so later I simply slowed myself down and worked on some retro things.
02
PART
NVIDIA NIM: Jensen Huang is really a good person!
FREE INFERENCE · OPEN MODELS
What NIM is
NVIDIA has packaged a batch of open-source large models into OpenAI-compatible APIs that developers can call for free, hosted on build.nvidia.com.
There are many models, with no limit on token quantity, but token rate is limited.
PS: using less popular, older models is often much faster
How to register
Go to build.nvidia.com, log in with an NVIDIA account, then go to the 'API Key' page and Generate one yourself. Save that key — you will need it next.
Integration method — direct connection
For an agent as advanced as opencode, you just fill in the key and use it.
The ones I use most are GLM-5.2 and deepseek. For minimax I actually use version 2.7 most — as everyone knows, versions after M3 became too slow and dawdling. Versions before 2.5/2.7 are genuinely stupid-fast; apart from being less clever, they are extremely convenient for small jobs, like raising horses and raising shrimp, though I have never found a use case for raising horses and shrimp. kimi2.6 was recently taken down, presumably in preparation for k3.
The second integration method — CC-Switch
CC-Switch is a 'configuration central manager' spanning a bunch of programming CLI tools: Claude Code, Codex, OpenCode, Claude Desktop, Gemini CLI, Grok Build, OpenClaw and Hermes. It is itself a Tauri desktop app, and the official site is ccswitch.io.
For claude code and codex you need the cc-switch plugin. I previously used another cli-proxy-api plugin for format conversion; cc-switch may have built it in after an upgrade, but I am not sure. I have not used it for a while.
For codex the goal-directedness is stronger — a single task often runs for an hour. cc is more divergent: if the prompt or some harness work is not done well, it may rewrite things beyond recognition. I have had to roll back code a few times; using it well carries some learning cost.
I am not especially religious about which tool is stronger. More often I just go by the feel of running the jobs I need to run. I have used TRAE a few times recently and it is much better than before, but because of the quota problem I did not make it my main tool.
✦ My own usage — in CC-Switch I configure NIM's GLM / DeepSeek / MiniMax for both Claude Code and codex, so I can vibe coding on both sides at zero token cost, and the context is not shrunk.
03
PART
Agnes: a free AIGC three-piece kit
TEXT · IMAGE · VIDEO
Image generation, video generation, text
Agnes is a large-model platform from a Singapore company, focused on generation across text, image and video, offering a free flash version in all three directions.
It provides a web version called pavo at: https://app.pavo-ai.work/
How good is it really
Many bloggers rave about how good this free model is — I do not know whether they are all taking sponsorship — but since it is free, we cannot say it is bad. Specifically how good it is, I ran some real tests.
I originally integrated this free API straight into a skill, letting the AI derive prompts from context and generate images directly, but the results were really not usable, so I removed it. For doing this at scale you still have to go with gptimage2.
Image generation
Generating an image of a girl in a maid outfit dancing a hot dance
Generating an image of a classical beauty holding a pipa, with snow in the background
A green, sunlit street with trees, a light-coloured wall and street fixtures, full of a bright spring feel, warm sunlight casting a halo, with light and shadow on the strands of hair and on the face. The character is a beautiful 18-year-old model, long hair falling naturally over the shoulders, one hand making a peace sign, the other hanging naturally, the face undeformed, the image fresh and natural and high-definition, a joyful atmosphere, extreme detail. Aspect ratio 9:16
Young-at-heart web-style photo shoot, a stunning fair-skinned girl, clean bare makeup, outdoor setting, low-angle shot, face looking straight at the camera, blue sky and white clouds, a pure and seductive feel, soft fair skin, skin smooth like jade, a sky-blue lazy knit cardigan, an atmospheric feel, a lazy posture, and the overall posture conveying a unique charm immersed in her own style, showing off a soft and cute temperament, high attractiveness, a 19-year-old girl with a pointed chin and balanced three-court-and-five-eye proportions, a petite frame, a shy expression, delicate porcelain-like dewy skin, on the cooler fair side, random makeup, random hairstyle and colour, random accessories, random actions, side lighting, hair strands lightly floating and tousled, shot from a random angle, random pose, with a lot of overall uncertainty, yet still good-looking
Create a high-resolution square (1:1) food poster featuring (your dish). Capture the dish with dynamic, explosive composition that seems to burst toward the viewer. Show realistic, mouth-watering texture and vivid colour. Place it against a deep black background, with professional dramatic studio lighting, varied depth of field and high contrast. Emphasise its premium, luxurious and visually striking qualities. Exclude any hands, people, text, logos or distracting tableware.
Style: photorealistic CG sci-fi horror film still, dark post-apocalyptic 'Sin City: 0' aesthetic, highly faithful to the original's monstrous texture and battlefield atmosphere. Medium: cinematic CG rendering, 4K ultra-high resolution, sharp detail, cold severe lighting, no noise. Main monster: the Oni滑瓢 in its ultimate corpse-demon form. Appearance and material: dark red muscle with exposed silver-white bone, skin dried and torn, muscle texture and bone outlines clearly visible, surface covered in moist bloodstains and wear marks. Limb features: multiple huge bone spikes growing from the back, grey-white and bone-textured, sharp-edged and covered in cracks; curved bone blades extend from the elbows, grotesquely shaped; the head is a dried demonic face with deep eye sockets, protruding facial bone, a cold expression full of oppression. Dynamic pose: the body leans slightly forward, the right arm extended, fingertip precisely pointing at a human character's forehead, arrogant and full of menace; the left arm hangs naturally, holding a severed head, fingers like withered claws, joints twisted, the overall posture demonstrating absolute strength. Human character: East Asian male warrior. Hair and face: neat short black hair, strands hugging the scalp, subtle wounds and blood on the face, resolute eyes, looking up at the monster, expression tense, showing a determination to fight. Clothing and accessories: a black skin-tight combat suit of matte rubber and metal panels, with slight wear and scratches, silver metal buckles and tubing on the neck and shoulders, sci-fi combat texture, battle stains on the face. Dynamic pose: the body leans slightly back, tense, arms hanging naturally at the sides, muscle lines faintly visible, showing vigilance and defiance when facing a giant monster. Environment and setting. Background: nighttime ruins of a city skyline, skyscrapers damaged and burning, thick smoke, neon signs and Japanese lettering visible in the distance, lights flickering hazily through the smoke. Mid-ground: a broken metal railing running across the frame, ground covered in rubble and ruins, scattered sparks and abandoned vehicles visible on the distant street. Atmosphere: the oppression and despair of an apocalyptic battlefield, the giant monster and the tiny human forming a strong visual contrast that highlights the disparity in power. Lighting and craft. Light source: ambient light of orange firelight and cool neon, the key light hitting the monster and the human from the upper side, forming high-contrast hard shadows that emphasise the solidity of muscle and bone. Texture: the wetness of muscle, the roughness of bone, the rubbery texture of the combat suit and the translucency of smoke are all highly reproduced, with clear material separation
In an Alpine environment, generate a high-end fashion winter editorial portrait with cinematic luxury aesthetics. The subject is a young woman sitting on a carved block of ice, legs naturally apart, one hand in a winter glove resting lightly against her cheek, her expression relaxed yet strong. Outfit: an oversized black down jacket on top, white wide-leg ski pants below, white luxury winter boots with red laces, and a plush white fur hat (covering the ears), with winter gloves on the hands. The setting is a broad snow-mountain landscape, with smooth undulating snow drifts in the foreground and majestic snow peaks in the background. The subject sits on the surface of a carved block of ice, and the sky is a gloomy winter sky. Lighting uses overcast natural light, overall soft and diffuse, with the light source coming from the gloomy sky. Shooting uses an 85mm lens with shallow depth of field, the subject in sharp focus, in fashion-editorial photography style. The composition is minimal, the subject centred, showing a cinematic frame, overall an ultra-clean luxury fashion editorial style. The palette is dominated by cool blue, white, soft grey and deep black, and the atmosphere is full of cinematic quality, high-end texture, serenity and power. The image pursues extreme realism, faithfully showing fabric detail and highly realistic ice texture, at ultra-high resolution with photo-grade realism, sharp focus and clear textural detail.
Visual style: an artistic style combining digital illustration with beautiful landscape painting, rendered with delicate soft brushwork and layered colour, overall showing a romantic, dreamy, cinematic light-and-shadow quality, with a calm and poetic atmosphere. Core elements: a vast coastal sunset landscape, blooming red and pink rose bushes in the foreground, a narrow rocky land bridge connecting to an island in the mid-ground, and a white lighthouse with dark coniferous forest standing on green hills in the distance, with a wide calm sea and a sky interwoven with brilliant sunset clouds. Specific content: the bottom of the frame is occupied by dense rose bushes, red and pink petals layered upon layered with clearly visible texture, dark green leaves interspersed among them, stems upright and natural. A rocky land bridge covered in low flowers and grass winds from the lower left to a small circular island at the centre of the frame, where vegetation is dense and dotted with small wildflowers of various colours, with a simple white lighthouse standing at the top, its tower straight, surrounded by a few tall dark cypress trees. The distant sea shimmers, and the afterglow of the sunset lays a bright golden path across the water, gentle waves lapping the rocks and forming delicate white foam. The sky shows a warm gradient from orange-red to pale purple, with thin soft clouds and warm diffuse light. Composition: vertical format with depth-perspective composition, the foreground flowers acting as a visual frame leading the eye toward the mid-ground bridge, which forms a natural diagonal leading line drawing the gaze to the distant lighthouse, with the horizon in the upper third of the frame, following the rule of thirds. Overall the spatial layers are clear, the foreground sharp and delicate, the mid-ground transitioning naturally, and the distance slightly hazed for aerial perspective to enhance depth. Rendering parameters set to high-precision digital drawing, using a wide-gamut colour mapping, moderate surface reflectance, rock showing rough granular texture, water using smooth gradients with specular reflection overlay, light coming from a low angle in the upper right forming soft rim-backlight contours and warm volumetric light, overall palette dominated by complementary warm orange and cyan-blue, creating a tranquil and beautiful dusk seascape.
Visual style: hard sci-fi cinematic CG rendering, hyperrealistic architectural concept art, cool and restrained colour, with silver-grey, pure white and pale blue as the main palette, an overall grand and solemn atmosphere, with strong futuristic industrial aesthetics and post-apocalyptic desolation, extremely crisp image quality and varied detail. Core elements: a gigantic futurist mechanical facility attached to and cantilevered from the edge of a towering white cliff, with a vast wasteland and a distant futuristic city skyline behind. The main structure consists of multi-level silver-grey metal platforms, dark alloy piping, vertical support structures and precision mechanical components, showing a heavy industrial texture and a high-tech feel. The cliff surface is white as snow or chalk rock, with pronounced vertical erosion textures and layered structure. The distant city has towers in dense rows, their outlines slightly hazy in aerial perspective, and the sky shows a pale, high-key blue-white. Specific content: the left of the foreground is a vast cantilevered building complex with complex structure, including cylindrical pipes, rectangular modules and grid-like metal railings. The building edges are cleanly cut, creating a strong contrast between the artificial and the natural against the cliff. Some heavy mechanical devices hang on the outer side of the cliff, cutting into the rock, showing the complexity of the engineering structure. The abyss below the cliff is bottomless, enhancing the sense of height. The city buildings in the background are densely arranged, with spires and streamlined towers, suggesting a highly developed civilisational ruin or a frontier colony. Light falls evenly, with soft highlights reflecting off metal surfaces and cool grey tones in the rock shadows. Composition: vertical composition from a high, downward-looking angle. The main building sits at the left third of the frame, with the line of sight running along the cliff edge toward the right rear, leading to the distant city centre. The foreground mechanical structures and the mid-ground vertical cliff form tight spatial layers, and the distant skyline provides environmental background, enhancing the sense of grand scale. Craft parameters set to full-frame sensor, using a 24mm wide-angle lens at f/8, shutter speed 1/500s, ISO 100, daylight white balance, focus mode set to hyperfocal, ensuring everything from foreground mechanical detail to the distant city outline stays extremely sharp, with sufficient depth of field, no distortion and naturally smooth light transitions.
Summary: it is quite prompt-dependent — with thin prompts you are likely to get extra limbs and fingers and similar defects; with identical prompts its performance is well short of local open-source models, and far behind closed-source ones.
Video generation
Image-to-video: convert to video, free rein
Using the image just generated
Image-to-video: convert to video, free rein — the girl in the image gracefully drinks and composes a poem.
Here I use a sample image from the official homepage
Image-to-video: convert to video, free rein
The model's comprehension is not great, so I removed the prompt guidance and let the model do it itself, which was indeed a bit better.
Summary: it probably cannot beat local LTX2.3, and it cannot be fine-tuned; its prompt comprehension needs improvement.
Text
I will not test this — you can just use GLM5.2 or deepseek, there is no need to bring it into the comparison.
///
LAST
Next step: running real business
NEXT
With these three freeloading tools in hand, many business scenarios are already covered. I will probably share some experiences from actual business applications later.
Freeload first, then pay; prove it works first, then scale up