Neural ExpeditionSubscribe
TEST RECORD · ← all field reports

FIELD REPORTS/IMAGE/ISSUE #87

Copy a pose from any photo onto your character

Posing a character in an image editor is still oddly painful: you either wrestle a skeleton rig in Blender, feed an openpose stick figure to a ControlNet, and hope the depth comes out right, or you retype a prompt twenty times. A LoRA maker who goes by lilylilith took a different route: show the model a photo of the pose you want, and let it do the rest.

MODELAnyPose
PUBLISHEDAugust 10, 2026
READ TIME3 min
TESTED BYNeural Expedition
CATEGORYIMAGE

Field notes

01What it does

AnyPose is a pair of LoRA weights for Qwen Image Edit 2511, the open image editor. You load them, give the model two images, your character and a pose reference, plus a ready-made prompt from the README, and it redraws your character in the reference pose: arms, legs, head tilt, even eye gaze. No stick figures, no ControlNet, no rig. It was built to work with the four-step Lightning LoRA, so generations come back fast, and the weights are Apache licensed.

The clever part is how it handles what it cannot see. If your character is framed from the waist up and the pose needs legs, it fills in the missing half plausibly, and a short prompt addition like what they are wearing steers that fill. The maker trained it on 3D characters posed in Blender, deliberately choosing positions the base editor could not understand on its own: handstands, bridges, deep bends.

02How to try it

The community demo Space by linoyts is the quickest test: upload a character image and a pose photo, and use the trigger prompt from the model page. Locally it runs anywhere Qwen Image Edit 2511 runs, ComfyUI included: load the Lightning LoRA plus both AnyPose files at 0.7 strength and paste the same prompt. The README on Hugging Face walks through the exact setup and the fix-it prompt lines for backgrounds and clothing.

03Caveat

It was trained on 3D-rendered people, so flat art styles like pixel art and cel shading often fail outright, and hyper-technical poses or scenes with more than one person trip it up. The scene can also swap unexpectedly: sometimes your character inherits the pose photo's background, which the README fixes with one appended prompt line. A V2 addressing these is planned. Note the demo Space is community-run, so it may change or rate-limit.

04What you can do with it

  • Build a character sheet: the same character in a dozen poses from photo references.
  • Redraw a mascot or game character in an action pose you found online.
  • Repose people between storyboard or comic panels without redrawing them.
  • Mock up fashion or product shots in a specific stance without a reshoot.
  • Fix a generated image where the pose came out wrong instead of rerolling.

Try the demo

View model page

Read this issue on neuralexpedition.com →