Quick Answer: A pose reference library is a searchable collection of images that shows how a human body holds a position, usually from several camera angles. RuiC-PoseAtlas is a free, static example with 288 pose groups shot from five angles (1,439 images) and 1,407 outfit sets. Use a pose to lock the skeleton and camera, then an outfit to set the look.
Most artists who go looking for a pose reference library end up with a folder of forty random photos, half of them cropped at the knees and none of them matching. You can't compare proportions across shots that differ in lens, height, and lighting. The fix isn't a bigger folder. It's a library built around consistent units. RuiC-PoseAtlas, published by HRuiCcc on GitHub, takes that approach with two atlases, and it's a useful case study for anyone building or choosing a reference system.
Why Organized Pose Libraries Beat Random Image Searches
A random image search answers "show me a pose like this." It doesn't answer "show me this pose from below" or "what does the far arm do when the torso turns." Those are the questions that actually block a drawing or a generation.
Organizing by pose group instead of by file solves that. Each group bundles the same body position across several viewpoints, so you can flip between them without losing the thread. This is the core idea behind the project. Its README describes the grid as one card per pose, with a thumbnail strip underneath for quick angle previews.
Here's the practical test: pick a pose, then ask whether you can see the same posture from the front, side, above, and below in under ten seconds. If you can't, your library is a pile, not a reference.
What RuiC-PoseAtlas Actually Contains
The site has two halves that share one design system and one language switcher. The pose atlas holds 288 groups and 1,439 images, with four to five angles per group: front, three-quarter, side, high angle, and low angle. Pose originals are 900 × 1200 PNG, and the grid loads 300 × 400 JPEG thumbnails.
The costume atlas holds 1,407 outfit sets, each shot as a person-free garment image in a 2:3 portrait format. These are split into eight seasonal and themed categories and tagged with 86 labels, each written in Chinese, English, and Japanese. The README reports about eight to ten tags per set, which is what makes search useful.
Two details matter more than the counts:
- Search covers every language. A query for "lace," "レース," or "蕾丝" hits the same outfit.
- Full-size images load on demand. Grids only pull thumbnails, so the page stays usable even with over a thousand entries.
Why Five Angles Per Pose Matter More Than You Think
Here's the counter-intuitive part. Most reference advice says to find one great image and study it. For structure work, that's often the wrong approach.
A single angle hides foreshortening. A pelvis tilted toward the camera looks shorter than it is, and a single front view can't tell you whether the shoulder is rolling forward or back. When you compare the same pose across the front, side, high, and low views, the inconsistencies become obvious. A limb that seemed fine in one shot clearly breaks in another.
This is also why multi-angle pose reference sets work well for AI-assisted work. When you feed a model a single pose image, it has to guess the depth. When you give it a matched set, you can pick the view that matches your camera and check the others for anatomy. The practical rule: if you can only save one image from a pose, you're making your later decisions blind.
Here's where most people stop and get it wrong. They save the image that looks best and skip the others. Keep the whole group.
Face-Free Mannequin Line Art: A Feature, Not a Gap
Many artists assume faceless mannequin images are a limitation. The README describes the pose images as face-free gray mannequin line art, and that's a deliberate choice worth understanding.
Faces pull attention. When a reference shows a real person's features, your eye tracks identity instead of structure, and generative tools can carry that identity into the output. A neutral mannequin keeps focus on the things you're actually studying: block volumes, joint placement, and the balance of the body under the pose.
There's a trade-off, and you should know it. Mannequin line art won't tell you how a real body compresses under fabric or how a specific skin texture catches light. For that, you'll still need photographic references or your own life drawing. Use the mannequin for skeleton and camera, and reach for other sources when you need surface detail.
Pairing Poses With Outfits for Character Design
The two atlases are built to be used together, and the workflow is simple. Here's a sequence that works without much fuss:
- Choose a pose group first, based on the gesture and camera height you need.
- Pick the angle that matches your final framing, then check the other angles to confirm the anatomy holds.
- Search the costume atlas by tag, not by number. Start with a broad term like "dress" or a colour, then narrow with a second tag.
- Keep the outfit reference on a separate layer or in a separate window, so the garment doesn't pull your pose decisions.
- Save both references together in a project folder, so the pairing survives your next session.
Pose sets the structure. The outfit then tells the viewer what the character is. The reason to separate them is practical: when you change the outfit, the pose stays valid, and when you change the pose, the outfit still fits the character. A single combined image can't do that.
The README suggests the costume images can be applied to existing characters without cutting out a person, because the garments are shot without a body. That's useful, but it depends on how closely the garment's cut and perspective match your character's. Expect to adjust.
Comparing the Two Atlases
Both halves share a layout, but they answer different questions. The table below summarizes the differences using the figures from the README.
| Dimension | Pose Atlas | Costume Atlas |
|---|---|---|
| Entries | 288 pose groups | 1,407 outfit sets |
| Images | 1,439 (4–5 angles per group) | 1,407 (one per set) |
| Full-size format | 900 × 1200 PNG | 2:3 WEBP, mostly 1366 × 2048 |
| Primary filters | Series → category → angle | Season → category → colour |
| Search fields | Name, number, category, angle | Number, English keyword, tags in three languages |
The contrast is the point. The pose atlas is about consistency across views, and the costume atlas is about searchable detail. If you're choosing a library for yourself, decide which of those problems costs you more time.
Caveats Before You Build on This
A few limitations deserve clear attention before you commit a workflow to any reference library.
Image rights come first. The code is MIT-licensed, but the README makes clear that the images are subject to their own usage terms, and it asks contributors to confirm their rights to any images they submit. Check the license for each image before using it in anything you sell or publish commercially.
The content is for adults. The project includes a notice that some material is for adult creators and isn't suitable for minors. Follow the laws where you live and the rules of any platform you publish to.
The repository is large. The README lists about 810 MB for the full image set, and the README's own estimate for the image folders is 654 MB for pose PNGs and 117 MB for costume WEBPs. Expect a slow first clone. If you only need code, the README describes a sparse checkout that skips the image folders, and you can skip the pose image check with SKIP_POSE_IMAGES=1 npm run verify.
Run it over HTTP. The pages load JSON with fetch(), which browsers block under file://. Start a local server, as the README shows with python3 -m http.server, instead of double-clicking index.html.
Frequently Asked Questions
What is a pose reference library?
A pose reference library is an organized set of images showing human body positions, usually from several camera angles. Good ones group the same pose across views so you can check proportions and foreshortening. RuiC-PoseAtlas is one free example, with 288 pose groups and 1,439 images.
How do you use pose references for AI art?
Start by choosing a pose group that matches your framing, then use the angle closest to your camera. Check the other angles for anatomy before you generate, because a single view can hide errors. Keep the pose reference separate from any outfit or character reference so each one controls one thing.
Why are the pose images faceless?
Faceless mannequin line art keeps your attention on body structure rather than identity. It also avoids carrying a real person's features into your output. The trade-off is that it won't show skin texture or how fabric drapes on a real body, so pair it with photographic references when you need that detail.
Can I use outfit references on my own characters?
Yes, the costume atlas is designed for that, since its garments are shot without a body. Match the cut, perspective, and proportions to your character before you commit. Also check the image's usage terms, because the repository says images carry their own rights.
Is a pose reference library worth building yourself?
It depends on how often you reuse references. If you return to the same poses, a curated library saves time. If you only need one reference per project, a quick search may be enough. Start with a small set of consistent multi-angle groups, not a large folder of unrelated images.
Your next pose session doesn't need a bigger folder. Pick one pose group from a pose reference library, save all five angles, and draw or generate from the one that matches your camera. Then pair it with a single outfit from the costume atlas and note whether the proportions hold. Pass this to a friend who keeps losing good references in a messy downloads folder.