The fourth to enter, two days late: it turns someone else's free image library into one-click image generation. 8 days to 700 stars

At the end of October, free pose and costume galleries by a Japanese AI illustrator blew up in the Chinese-speaking world, and a dozen-plus "Chinese versions" raced to launch. open-pose-costume came two days after the first, and most of its material isn't its own. It still picked up roughly 700 stars by adding one more layer, "pick a pose and outfit and generate directly," plus the author posting daily and Taiwanese tech media reprinting it. It shows that inside a wave already crowded with entries, a latecomer can still claim a slice through productization, but the slice won't be big.

An article from a Taiwanese tech outlet

11 月 4 日下午 3 点半(台北时间),台湾科技站電腦王阿達发了一篇稿子,标题是《Open Pose & Costume 收錄超過 5,800 張參考圖,人物、姿勢、服裝都有,讓你 AI 生圖更容易》,同时推到 X 和 Threads。稿子开头写得很实在:现在很多人用 AI 生成人物图,为了控制动作和衣服都会上传参考图,「不過多多少少還是會碰到你手邊剛好沒有適合的參考圖」。

A day earlier, another Taiwanese tool site, DeTools, had already covered it. Over the next two days, Yahoo News Taiwan and Hong Kong's ezone picked it up. On Trendshift's numbers, the repo's stars spiked at a single point across the whole curve. 11-03 (UTC) alone added 300, and that day it reached #15 on Trendshift's TypeScript chart.

The repo was built by hongye007, Kevin Zhou from Shanghai. Before this, his GitHub was almost entirely forks with 0 stars each, and his X account @HongyeMain had only 331 followers. He created the repo late at night in October 29, by which point the first entry in the same wave had been live for two days and had thousands of stars.

What it is

A bilingual Chinese-English site (live at omilic.com, with a GitHub Pages mirror), split into two halves:

Half one is the reference gallery. The headline numbers in the README are 312 pose sets (1,559 angles), 2,769 outfits and 1,527 characters, for 5,855 images total. Poses are gray mannequins, each set shown from five angles (front, three-quarter, side, top, bottom), many with depth maps. Outfits come in realistic and anime styles, filtered by 8 categories such as season and party. You can search, favorite, download originals, and share links that preserve your filters.

Half two is the one-click generation workspace. Pose and outfit are required, character reference is optional. In the top right, the user enters their own API base URL and key. The default is an OpenAI-compatible relay, atomflow.vip, with model gpt-image-2. Two or three reference images are sent together to generate "this person, in this outfit, in this pose." You can also upload a photo of an outfit and "extract clothing" as a reference. Results can be shown side by side with the original, or overlaid in one frame with a drag slider for comparison. The last 20 results are kept in the browser. A companion repo, open-pose-costume-skill (33 stars), packages the same flow as a skill you can call in natural language inside Codex.

Where the material comes from is spelled out in SOURCES.md. The poses, realistic outfits, anime outfits and depth maps all come from four free sites run by the Japanese illustrator ギア (@giaaibusi). Portraits come from another Japanese AI portrait gallery. The Chinese pose names are taken from the top entry, anhao/ai-pose-gallery-zh. The author's own contribution is only 42 outfits and 24 pose sets (AI-redrawn or extracted), and that part is explicitly marked as not for commercial use.

How the stars grew

(GitHub daily star / fork counts, Beijing-time snapshots; 11-06 during the day the GitHub page already showed 717 stars and 185 forks)

After the spike it fell quickly. 11-06 added only 26 in a day. Forks are a quarter of the stars, a ratio normal projects never reach. Many people are forking it, changing the domain, and deploying their own copy.

Where people came from, by share:

1. A spillover from the wave (inferred, high confidence). The demand comes from the #1 repo, anhao's library: 10-29 it gained 1,074 stars in a single day, and GitHubDaily, Zhihu and CSDN all wrote about it. This repo was created the same day and got listed on Chinese navigation sites alongside a dozen similar ones.
2. Taiwanese and Hong Kong tech media (confirmed). DeTools 11-03, 電腦王阿達 11-04, Yahoo News Taiwan and ezone 11-05. The timing lines up with the spike. Their headlines all landed on the same angle: "5,800 reference images + one-click generation." The mainland accounts that covered it wrote about the top entry and didn't mention this one.
3. The author posts daily (confirmed). 10-30 posts "Hidden gem: open-source free projects, every pose you want is here" on V2EX and on Weibo. On X, 10-31 posts two tweets, 11-02 "just had another big update," 11-03 announces the switch to the domain omilic.com, and from then on posts daily, each with the repo and a live demo link. YouTube Shorts has at least two posts, "three references, generate a new look in one click." Someone also shared a Traditional Chinese version in a Facebook group. 8 days, from v1.0 to v2.1, each release comes with fresh material, so the media always has something new to write about. But with 331 followers, the traffic from self-promotion alone won't be large.

No sign of it on Hacker News, Reddit or Product Hunt, and no major account tagged it for a repost. I couldn't fully pin down which single push produced the 11-03 spike of +300. It was likely media coverage, group reposts, chart exposure and that day's self-promotion stacking together.

Who it's competing with

The same Japanese material spawned over a dozen Chinese repos after 27 in October:

Its feature set is one of the most complete in this batch, yet it has about a ninth of the top entry's stars. The gap isn't features, it's timing. The top entry appeared first, got cited by navigation sites, tutorial sites and big accounts, and everyone who came later saw it first. People are already copying it: 11-06, a newly created brian479974/muse, states it is "adapted from open-pose-costume," and others have bundled originals from several repos into releases.

Why it still got a slice

• It rode the demand. Once models like gpt-image-2 and Nano Banana, which accept several reference images, became common, "one pose image + one outfit image" became the standard way to generate character images with AI. Free, commercially usable reference images were in short supply.
• It did one more layer than the others. Most copies are just "Chinese gallery sites": you look at the pictures and then go elsewhere to generate. This one wired the gallery straight into generation, lets users bring their own key, runs a relay service online, and works on phones. Media were willing to write about exactly that layer.
• It kept updating and kept promoting. Each version came with an image and a post, giving reposters something to share.

But these differences only determine how big a slice it got in this wave. They don't determine whether there was a wave at all. The material belongs to someone else, the part it built itself is small, and others in the same batch are already building that generation layer.

The money question

It's an open-source project and makes no money. One note: the project is entirely free, generation defaults to atomflow.vip, and the README links a Feishu "free trial guide." Whether the author has any partnership or referral arrangement with that relay isn't visible from outside (unverified).

The licensing is the bigger issue: the original ギア site bars redistributing its images, and the author says 10-30 got "extra permission" from the original creator to put about 1.95 GB of source images in the repo. But SOURCES.md itself says it "has not been independently verified." Japanese outlets are already looking into the licensing of these "open source" libraries.

If it were you

C: This wave is over. What's worth learning is how latecomers claim a slice. The same material already has a dozen-plus Chinese repos. The top one, with 6,000 stars, holds the entry point, and its own daily gains have dropped to twenty or thirty. Building another copy now won't earn many stars, and it carries licensing risk.

Who it suits: people who can build a static site plus a generation workspace that calls an image API within two or three days, who already hang out in Japanese and Korean AI image circles, and who want to use one project to build stars and followers.

Who it doesn't suit: people chasing these pose libraries right now; people who want to build a project to maintain over the long run and grow slowly (this kind of repo's curve is one spike followed by decline); people who are sensitive to copyright risk.

How to do it:

• Watch the sources. When "free for use" or "commercial use OK" assets appear on Japanese or Korean X, first judge whether a large group of people who don't read the language would want them.
• If you can't get there first, don't build a second "translated version." Build the next layer instead: connect it to generation (like this one does), make it a ComfyUI node or agent skill, or make a workspace that works on phones. Media want something new to write about. A second "Chinese version" is no longer news.
• Sort out licensing first. For material whose original site bans redistribution, build an index that links back to the source rather than moving the originals into your repo. Getting a public license from the original creator is itself something latecomers can't copy.
• Every release needs something new you can screenshot, paired with a post.

How to get stars: leave the link under the original post and the Chinese reposts the day it launches. If the big Chinese accounts already have the top spot, try the tech tool sites in Taiwan and Hong Kong (DeTools, 電腦王阿達 and similar). They have low bars for submissions and long repost chains, and this repo's peak came from them.

Deadline: for each wave like this, the top spot is only open for about 24 hours after the original post. The "next layer" spot lasts roughly two or three more days. This wave has passed; the next one would be a new repo from Gear or other Japanese/Korean creators.

How to tell if we misjudged it: ① If this repo is still gaining hundreds of stars a day two weeks later, then "reference images + one-click generation" is a lasting demand, and a version with your own assets and your own licensing is worth building. ② If the original author demands takedowns and these repos disappear, then scraping assets is a dead end and only indexes and original assets are left. ③ If someone builds an original-asset version of the same tool and gets a thousand stars in a week, then readers are buying clean assets, not feature count.