GALIAIS-Nodes 05 姿势动作

A character standing stiffly in a good outfit is a mannequin, not an image. The Pose & Action node (05 姿势动作) is the layer that gives the character something to do - posture, gestures, motion, and how they relate to whoever else is in the frame. It's also, quietly, the layer with the largest and most specific Danbooru tag vocabulary, which means it's where using the right tag name matters most.

What you fill in

Four content fields plus a supplement, each with a 使用 toggle:

• 整体姿态 (overall posture) - standing, sitting, lying, leaning, the body-level pose.
• 肢体手势 (limbs/gestures) - arms crossed, hand on chin, thumbs up, victory sign, the detailed limb positions.
• 动作行为 (action/behavior) - running, jumping, reaching, turning around, motion with intent.
• 互动关系 (interaction) - the relational tags: holding hands, hugging, looking at another, hand on shoulder. This is the field that makes a two-character scene read as two characters.
• 姿势补充 (pose supplement) - free-text extras, multiline.

Plus the shared controls: 角色槽位 (全局 / 角色1–8) and 角色标签 for multi-character grouping, 启用 and 权重, and the random/AI block (启用随机Tag, Tag生成模式, 随机种子, optional AI服务商 for the AI-collaborative modes), with optional DB, 上游上下文, and 上游提示词段.

Outputs

The standard trio: 姿势段 (section object → composer), 文本 (plain text), 元信息JSON (contribution + random audit).

Where it earns its keep

The 互动关系 field is the star. Prompting 2girls with no interaction tags gives you two disconnected figures that the model often merges awkwardly; one interaction tag (hugging) makes the sampler compose them as a scene. This is also the layer where the pack's multi-character mode really pays off - tag the pose node to 角色1 and the interaction reads as that character's action, and the composer keeps per-character grouping clean.

The one genuine trap: pose tags conflict all the time, and the composer only dedupes exact strings. sitting in this node and standing in a character tag will both reach the sampler, and the model will fight itself. This is exactly the case the pack's AI Conflict Resolver was built for, but for everyday use, keep one pose authority per character and check 文本 before you run.

The honest read

With pose, the tag-vocabulary advice from the community applies harder than anywhere: models trained on Danbooru understand the real pose tags (one knee up, hands behind back, contrapposto) far better than prose descriptions. Look up the tag before you hand-wave. And 肢体手势 is the field that fixes the eternally "stiff hands" look - if every render has hands fused to the body, that's this layer being empty.

Install

Pack-wide, once:

git clone https://github.com/GALIAIS/GALIAIS-Nodes.git ComfyUI/custom_nodes/GALIAIS-Nodes

Restart ComfyUI, or ComfyUI Manager → "GALIAIS-Nodes". No pip dependencies; the Danbooru DB is separate if you want dictionary-backed picking. New pack, no community writeups - lean on 元信息JSON and the README.