What Is AI Pose Gallery (Chinese Edition)? A 3-in-1 AI Art Reference Gallery for Poses, Costumes, and Portraits

Quick Answer

A Chinese-localized AI art reference gallery: pose, costume, portrait — 4,345 entries, 8,690 images, ComfyUI plugin.

The slowest part of AI art is often not generating the image but pinning down the references: what pose the character holds, what they wear, what they look like — three reference images to hunt down one by one. The three reference galleries published by Japanese author ギア (@giaaibusi) already solved that problem, but their metadata was in Japanese and spread across three separate sites. The Chinese edition, maintained by anhao, merges all three into one, translates everything into Chinese, and adds a composition studio that assembles prompts plus a native ComfyUI plugin: 1,439 poses, 1,407 costumes, and 1,499 portraits — 4,345 entries and 8,690 images in total. The repository was created on October 27, 2026, and had about 4,951 stars when verified on November 1, 2026. Two boundaries first: the gallery ships with safe mode on by default and is intended for compliant reference use only; all images remain the copyright of the original author, and repackaging them for resale is prohibited.\n\nDisclaimer: this site is not affiliated with anhao or the original author. This article is product-technology reporting and analysis, not usage, procurement, or licensing advice. Scale, star counts, and feature details follow the repository pages, and star counts change over time; copyright and license details follow the repository disclaimer and LICENSE.

1. What it is: three Japanese reference sites, rebuilt as one Chinese workbench

Start with the provenance. The original sites are three AI art reference galleries published by Japanese author ギア (@giaaibusi) — one each for poses, costumes, and portraits — with all metadata in Japanese. The Chinese edition, maintained by anhao, did three things in its v2.0: first, merge all three sites into a single repository and a single online demo; second, localize everything, translating the action system, pose categories, costume styles, outfit tags, portrait sections, hair lengths, and 51 hairstyles into Chinese, so auto-assembled prompts carry no Japanese residue; third, mirror all 8,690 images locally, thumbnails and full-resolution originals, for fully offline browsing and static hosting.\n\nThe numbers, re-checked against the README scale table: 1,439 pose entries with 2,878 images; 1,407 costume sets with 2,814 images; 1,499 portraits with 2,998 images (full-size 1600×2000 high-resolution WebP). 1,439 + 1,407 + 1,499 = 4,345 entries, and 2,878 + 2,814 + 2,998 = 8,690 images, matching the repository's headline of 4,345 fully Chinese reference entries and 8,690 local high-resolution images. The repository is about 1.1 GB with roughly 8,735 files, so leave disk room before cloning it locally.

2. What each library holds: poses for action, costumes for outfits, portraits for faces and hair

The pose library answers what the character is doing: 26 action categories, 288 pose groups, and 5 viewpoints (front, three-quarter, side, back, top-down), with two browsing modes — by pose group or image by image — and multi-angle switching inside each card. The practical point is choosing the viewpoint first: the same pose constrains a generation model very differently from the front, which fixes composition, versus the side or back, which fix body orientation and occlusion.\n\nThe costume library answers what the character wears: multiple outfit categories organized by season and occasion, with 86 Chinese detail tags and matching English prompt keywords. The Chinese tags matter for search — instead of translating a knit cardigan into English before searching, you search in Chinese directly, and the matching English keyword is already paired for the prompt.\n\nThe portrait library answers what the character looks like: 893 front-face references plus 606 styled portraits, 1,499 in total, with 4 hair lengths (short, medium, medium-long, long) and 51 hairstyles, plus curated collections and trending lists for one-click filtering. For character sheets, the value is not a single face but a comparable reference set — fix the section, narrow by hair length, and iterate a character's look far faster than flipping through images at random.\n\nAll three libraries share one interaction pattern: real-time Chinese keyword search, lightbox previews of full-resolution originals with download, and one-click add to a composition recipe. That last point is the key to the next section: the three libraries are not three separate albums but three slots of one recipe.

3. The composition studio: three reference images into one recipe, prompts assembled automatically

The composition studio (/studio) is the Chinese edition's biggest addition over the original sites. The flow is 3-in-1: pick one image from each library — portrait, costume, pose — to form a character recipe, or randomize a character in one click; beyond the three core slots, any number of extra reference images can be added by upload, drag and drop, clipboard paste, or image URL.\n\nOnce the recipe is set, the studio does three things. First, a composite preview and export: a 3-in-1 high-resolution reference collage (1800×1200, composed live on canvas) that can be downloaded as a PNG — the multi-image reference input for a generation model. Second, 12 style presets with purely Chinese structured prompts: from no preset (follow the references) through realistic portrait photography, Japanese-Korean magazine editorials, retro film photography, 2.5D game key art, Japanese 2D illustration, Shinkai-style animated film, black-and-white manga screentones, 3D next-generation character rendering, 3D designer-toy texture, cyberpunk neon, and character design sheets on pure white — prompts assembled automatically from the chosen materials. Third, direct connections to five AI image protocols: Google Gemini, OpenAI GPT Image, Alibaba Bailian Qwen, ByteDance Volcengine Ark Doubao, and local ComfyUI, with generation history persisted locally in the browser's IndexedDB, so switching pages never interrupts an in-flight generation.\n\nFor users who do not write code, the studio means selecting images is writing the prompt: where you once flipped through three sites and hand-wrote an English paragraph describing how three images relate, you now pick three images and a style, and a structured prompt is assembled. For workflow users, the next section's plugin is the main event.

4. The native ComfyUI plugin: pick images on the canvas, four nodes for three libraries

The repository doubles as a drop-in ComfyUI plugin with no extra Python dependencies, reusing the torch, numpy, Pillow, and aiohttp that ship with ComfyUI. Installation follows the README: clone the repository into ComfyUI/custom_nodes/ai-pose-gallery-zh, restart ComfyUI, then right-click an empty canvas area and open the AI character 3-in-1 reference gallery menu.\n\nThe plugin provides four nodes: AI pose reference, AI costume reference, AI portrait reference, and the AI character 3-in-1 composition recipe. A single node outputs its reference image as an IMAGE tensor plus the image ID and a description or prompt; the recipe node outputs the portrait, costume, and pose images, the 3-in-1 collage, and the auto-assembled Chinese and English composition prompts in one go. Nodes embed a visual picker covering all 4,345 images with Chinese labels, filterable by category, viewpoint, and hair length or searchable in Chinese, filling the node in one click with live thumbnail previews inside the node.\n\nBeyond plugin mode there is a web-to-local mode: if you pick images on the online demo or a local dev server and want your own machine's ComfyUI to generate, start local ComfyUI with cross-origin headers enabled (--enable-cors-header), switch the studio's API settings page to local ComfyUI, and test the connection (default http://127.0.0.1:8188). The collage or a single reference image can then be uploaded to local ComfyUI, injected into a custom API workflow, queued, and fetched back as a finished image, with support for importing and exporting generic .api.json workflow templates. The split is clear: plugin mode suits users who live in ComfyUI, web-to-local mode suits users who pick a recipe on the web and hand generation to their own machine.

5. Who it fits, and the boundaries: read the license and copyright terms first

Who benefits most. Tier one, direct fits: people doing AI illustration, character design, or comic storyboards — three libraries covering action, outfit, face, and hair in one place, with Chinese search removing the translate-then-search step. Tier two, ComfyUI users: plugin mode embeds image selection in the canvas, so recipes and prompts flow straight into workflows without exporting images back and forth. Tier three, agent and workflow developers: the repository's skills/ directory ships eight agent skill packs that let an AI agent retrieve the three libraries as references for image or video generation, using relative paths into the repository's img/ directory — symlink them rather than copying, and keep the whole repository intact.\n\nFour boundaries come first. First, copyright: all pose, costume, and portrait reference images come from the original author's public sites, and copyright remains with the original author; the repository is a Chinese translation, organization, and engineering study archive, and repackaging the gallery images for resale is prohibited — users bear sole legal responsibility for misuse. Second, compliance: the gallery ships with safe mode on by default and is intended for compliant reference use only; the repository disclaimer limits use to personal study, human-motion research, lawful artistic creation, and technical exchange, and prohibits unlawful uses and unauthorized real-person face swaps. Third, licensing: the repository license field reads Other (NOASSERTION), not a standard open license such as Apache or MIT — read the LICENSE file and the full disclaimer before any commercial use or redistribution. Fourth, scale and positioning: about 1.1 GB and 8,735 files, every file under 2 MB, within Cloudflare's free-tier limits; but this is a reference gallery and plugin, not a hosted generation service — its value is the reference data, Chinese search, and recipe assembly, not generation compute.

6. Sources and disclaimer

Sources (verified November 1, 2026):\n1) https://github.com/anhao/ai-pose-gallery-zh — repository page: created 2026-10-27T12:24:37Z, about 4,951 stars and 1,259 forks at verification time, license Other (NOASSERTION); README scale table: 1,439 poses / 2,878 images, 1,407 costumes / 2,814 images, 1,499 portraits / 2,998 images, totaling 4,345 entries / 8,690 images; original author ギア (@giaaibusi) and Chinese maintainer anhao; online demo and original-site links; full disclaimer and copyright statement;\n2) https://ai-pose-gallery.anhao.app/ — Chinese-edition online demo: merged three-library browsing and the composition studio entry;\n3) Repository README sections on core features and ComfyUI usage: 26 action categories, 5 viewpoints, 86 Chinese detail tags, 4 hair lengths and 51 hairstyles, 12 style presets, 5 AI image protocols, 4 ComfyUI nodes with a 4,345-image picker, and the two usage modes.\n\nDisclaimer: this site is not affiliated with anhao or the original author. This article is product-technology reporting and analysis, not usage, procurement, or licensing advice. Scale, star counts, and feature details follow the repository pages, and star counts change over time; copyright and license details follow the repository disclaimer and LICENSE and may change.

The three libraries at a glance (repository README scale table, verified November 1, 2026)

Library	Entries	Local images	Search dimensions
Pose	1,439	2,878	26 action categories, 5 viewpoints, 288 pose groups
Costume	1,407	2,814	86 Chinese detail tags plus English prompt keywords
Portrait	1,499	2,998	893 front-face references plus 606 styled portraits; 4 hair lengths, 51 hairstyles
Total	4,345	8,690	Fully Chinese-indexed libraries, with trending lists for one-click filtering

Entries per library

★ Pose1,439 entries

Costume1,407 sets

Portrait1,499 portraits

数据来源：Source: repository README scale table (verified November 1, 2026); 1,439 + 1,407 + 1,499 = 4,345

Frequently Asked Questions

How does the Chinese edition differ from the original sites?

Three ways: merged — the original pose, costume, and portrait sites were separate, while the Chinese edition keeps them in one repository and demo; fully localized — the action system, categories, tags, and hairstyles are all translated into Chinese, so auto-assembled prompts carry no Japanese residue; and extended — with a 3-in-1 composition studio and a native ComfyUI plugin, where the original sites were browse-only.

Can I use it without ComfyUI?

Yes. The online demo supports browsing all three libraries, picking images, and building recipes directly, and the composition studio connects to four cloud protocols — Gemini, GPT Image, Qwen, and Doubao — for generation. The ComfyUI plugin is a bonus for users who live in ComfyUI, not a requirement, and local ComfyUI can also take over generation through the web-to-local mode.

Can the gallery images be used commercially or resold as a pack?

No. The repository disclaimer states that all reference images come from the original author's public sites, copyright remains with the original author, and the repository is a Chinese translation, organization, and engineering study archive — repackaging the gallery images for resale is prohibited. Personal study and compliant creative reference are the stated uses; read the LICENSE and the full disclaimer before any commercial use or redistribution.

What does the gallery's default safe mode mean?

The repository states that the gallery ships with safe mode on by default and is intended for compliant reference use only, with the disclaimer limiting use to personal study, human-motion research, lawful artistic creation, and technical exchange. The applicable boundaries follow the repository disclaimer; this article covers only compliant uses — pose, costume, and portrait reference, the plugin, and prompt assembly.