DEV Community: yao zhong

Create a Custom Windows Cursor Pack with AI: A Step-by-Step Guide

yao zhong — Sat, 08 Aug 2026 09:42:13 +0000

Want a cursor set that actually matches your desktop instead of being a random mix of shapes? The usual options are limited: browse a cursor pack site for hours, or open an icon editor and try to redraw every state yourself. A text-to-cursor generator removes most of that work by building the whole set from a short description.

Why a coordinated pack matters

Windows assigns a different cursor for each state: the normal arrow, the text caret, the busy spinner, the resize handles, and several more. A good pack keeps all of those states visually consistent. When they do not match, the desktop feels broken even if every individual icon looks fine on its own.

What a text-to-cursor generator does

A tool like CursorHero turns a plain text description into a complete family of coordinated cursor states. You type something like "neon cyan with rounded corners" and the generator produces the whole set, not just one arrow. The output downloads as native .cur files along with an INF file, so applying the theme takes a few clicks in Windows pointer settings.

How to create your own pack

Describe the style you want. Be specific about color, shape, and mood, because the prompt drives the entire set.
Preview the result in the browser. Live preview helps you check each state against your current desktop theme before you commit.
Generate the pack. The tool builds all ten states as a coordinated family.
Install it. Download the .cur files with the INF file and apply them through the mouse settings page.

Tips for better results

Mention the accent color you want to match, for example the taskbar or window borders.
Say whether you use dark mode, since cursor contrast looks different on light and dark backgrounds.
Keep the description simple. Cursor shapes need to stay readable at small sizes, so avoid overly busy details.

Try it with a free preview

The CursorHero site offers three free previews so you can test the workflow before paying for a full pack download. Visit CursorHero and describe the style you have in mind for your Windows cursor pack.

Building a Two-Reference AI Portrait Pipeline with GPT Image 2

yao zhong — Fri, 07 Aug 2026 07:33:57 +0000

Building a Two-Reference AI Portrait Pipeline with GPT Image 2

Consumer AI photo products have a hard constraint that most demos ignore: the subject is not a single person, and the input is not a studio session. An AI engagement portrait needs to take two separate reference photos of two different people and produce one shared portrait that looks like both of them. This post walks through the pipeline design behind that kind of product, using a live example at the end.

The input problem

The first design decision is what the user must provide. A one-photo prompt is easy; two photos introduce identity consistency across two subjects. In practice the cleanest contract is one clear photo per person, with faces well lit, similar framing, and no heavy filters or extreme angles. The product then treats both images as fixed references and never merges them into a single averaged face.

The generation layer here is GPT Image 2. The two references are passed as image inputs, and the prompt supplies the scene, the relationship between the subjects, and the output constraints. The service offers eight editorial styles, including Korean Garden, Beach Romance, Enchanted Forest, Classic Elegance, Golden Hour, Bohemian Dream, Urban Chic, and Twilight Magic. Each style label is essentially a shorthand that anchors lighting, wardrobe mood, and backdrop in the generated scene.

The preview and credit model

A portrait generator needs a trust loop before any payment. The product's approach: after email verification, each account gets one watermarked preview at 1K resolution. That preview is the quality gate. If a provider or storage request fails, the attempt does not consume a credit, which matters when users iterate across styles with the same two photos.

Paid access is split between one-time HD packs and monthly plans. The one-time packs cover 20, 50, or 120 portraits, all HD and unwatermarked, with credits that never expire and failed renders refunded. Monthly plans start at the Basic tier and scale up to Pro, which also includes commercial-use rights per the published terms. Prices are time-sensitive and should be rechecked on the site.

The honesty constraint

The most important part of the pipeline is labeling. Outputs are AI-generated from the two reference photos, and the product states that they are not photographs from a real engagement session. That distinction matters legally and practically: usage rights are plan-dependent. Personal use covers the trial, Basic, and standard packs; commercial-use rights come with Pro, per the published terms.

A working example

If you want to see the two-reference workflow in production rather than in diagrams, the studio AI Engagement Photos is built on this exact pattern: two uploads, a style choice, a watermarked preview after email verification, and HD packs without expiry. The same honesty rules apply there, so it is a good reference implementation for the category.

Xiaomi MiMo models: a community guide to getting started

yao zhong — Fri, 07 Aug 2026 03:17:36 +0000

Xiaomi MiMo is an open-source, reasoning-first family of large language models developed by Xiaomi, with weights published through the official XiaomiMiMo organization on Hugging Face and source code on GitHub. If you are evaluating MiMo for coding, agents, or long-context workloads, the official docs at mimo.mi.com are the primary reference for model names, licenses, benchmarks, and API pricing.

One thing that takes time when starting out is finding the right resources: the Hugging Face organization hosts model cards and checkpoints, the GitHub organization hosts inference and tooling code, and the API documentation lives separately. A useful starting point is the Xiaomi MiMo AI community guide, an unofficial community site that collects tutorials, comparisons, and FAQ-style explanations, and links each claim back to official Xiaomi sources. The site is not affiliated with Xiaomi Inc., which makes it safe to read alongside the official documentation.

A practical workflow: start with the official model cards to pick a checkpoint, skim the community guide for setup notes and common pitfalls, then verify pricing and context limits against mimo.mi.com before building anything on top of it. This article is an independent community summary, not an official Xiaomi publication.