👘 Kisekae Pose

Most of this pack's sections are lists of nouns. Pose isn't - it's the one section with a free-text field sitting right next to its dropdowns, and that field is the difference between a character that renders and a character that renders doing something.

👘 Kisekae Pose has four fields: pose, hands, action, and prose. The first three are booru tags. prose is a multiline sentence slot.

How it works

Same as every section node: character in on KISEKAE_CHAR, this node's preset pick replaces the whole pose section if set, then per-field overrides apply with typed text beating the dropdown.

The interesting part is what happens at render time. The default template, anima-mixed, gives prose its own line at the bottom:

{style.quality}, {scene.composition}, {scene.lighting}
{triggers}, {identity}, {hair}, {head}, {body}
{outfit}
{style}, {pose}
{pose.prose}
{scene}, {head.expression}

That extra line exists because of how 2026-era anime models read prompts. Anima is a 2B Cosmos-Predict2 model wrapped around a Qwen3-0.6B encoder - it genuinely understands a sentence, and it understands Danbooru tags better than sentences. Mixing both, tags first and one natural-language line for the body language, is the layout the author shipped as the default for that reason. If you're on an SDXL-lineage model (Illustrious, NoobAI, Pony, WAI), switch the template to illustrious-tags, which is the same layout minus the prose line - those encoders will treat your nice sentence as a bag of half-recognised tokens.

The inputs that matter

pose covers the big postures - standing, sitting, kneeling, lying, on back, squatting, crossed legs, leaning forward. hands is the one to be careful with: v, peace sign, pointing at viewer, hand on own cheek, own hands clasped and so on. Hand and arm tags are still the most reliable way to get mangled fingers, so if a render goes wrong at the extremities, clear hands before you blame the checkpoint.

action gives you eating, reading, dancing, casting spell, fighting - the verb layer that pose tags don't cover. And prose is where you write the sentence the tags can't say: "she leans against the doorframe, one hand still on the handle, rain on her shoulders."

For an _append note: appending a pose tag to sitting gives you sitting, leaning back, which the model will usually reconcile fine. Appending two mutually exclusive postures gets you a coin flip. Replace, don't append, for those.

Output is one char.

Install

ComfyUI Manager → search kisekae (Registry: comfyui-kisekae), or:

cd ComfyUI/custom_nodes
git clone https://github.com/Kagejitsu/comfyui-kisekae

Restart ComfyUI. Zero Python dependencies - no wheels, no compilation, nothing to rebuild after a ComfyUI update. The pack's example workflow also wants LoRA Manager for its LoRA loader; the Kisekae nodes don't.

Common issues

My prose line is missing from the prompt. Either the selected template doesn't include {pose.prose} (check with the Templates page in Tansu, or just switch to anima-mixed), or something upstream cleared the field. Kisekae Debug Prompt shows a per-placeholder breakdown that answers this in one glance.

Prose is in the prompt but the render ignores it. On a tag-heavy SDXL anime model that's expected - see above, switch templates. On Anima, a sentence buried after forty tags competes with them; keep it to one line and keep the tags honest.

Pose and composition contradict each other. full body composition with sitting is fine; portrait with full body isn't. Composition lives in Kisekae Scene, and both land in the same prompt, so set them together.

Scene presets aren't applying pose. A section node's preset dropdown only lists presets that define that section. If you want examples/scenes/dramatic's pose, scene, lighting and expression all at once, use Kisekae Load Preset with mode: merge - that's exactly the case the mode exists for.