README.md

---
title: Krea 2 Pose ControlNet
emoji: 🎨
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 6.22.0
app_file: app.py
pinned: false
short_description: Generate characters in a pose from a reference image
python_version: "3.12"
models:
  - krea/Krea-2-Turbo
  - thedeoxen/Krea-2-pose-controlnet
---

# Krea 2 Turbo — Pose ControlNet LoRA

Drive generation with a body pose, either from a ready-made skeleton image or
from a normal photo whose pose you can edit interactively before generating.

## Two input modes

- **Pose Image**: upload an already-rendered OpenPose/DWPose skeleton image; it
  is used as the control input exactly as provided.
- **Regular Image**: upload a normal photo; a DWPose skeleton is auto-extracted
  and shown in an interactive editor where you can **drag individual joints** to
  tweak the pose. The (possibly edited) skeleton is what conditions generation.
  (Editor inspired by
  [linoyts/Flux-2-control-pose](https://huggingface.co/spaces/linoyts/Flux-2-control-pose).)

## How it works

1. **Pose extraction**: photos are processed with DWPose (via `controlnet_aux`)
   to produce an OpenPose skeleton map on a black background.
2. **Conditioning**: the pose map is fed to the Krea 2 model through the
   [Ostris Edit](https://huggingface.co/ostris/Krea2OstrisEdit) community
   pipeline — the pose image enters both the Qwen3-VL text encoder (as a vision
   token) and the transformer sequence (as a clean VAE reference latent at t=0).
   The pose image is scaled to the nearest **~1 MP preferred resolution bucket**
   (an exact replica of ComfyUI's `FluxKontextImageScale`: nearest-aspect
   bucket → center-crop → Lanczos), and that **same** scaled image drives both
   the VAE reference latents and the output canvas. This keeps the pose latent
   grid and the output grid identical *and* runs the model at its native ~1 MP
   resolution — which is what makes the model actually follow the control pose
   at full quality.
3. **Generation**: the model generates a new image following the body pose while
   the text prompt defines appearance, clothing, and scene.

## Model

- **Base**: [krea/Krea-2-Turbo](https://huggingface.co/krea/Krea-2-Turbo) (8-step
  distilled, `guidance_scale=0` disables guidance)
- **LoRA**: [thedeoxen/Krea-2-pose-controlnet](https://huggingface.co/thedeoxen/Krea-2-pose-controlnet)
  (recommended weight 0.8–1.0, ~10 steps, CFG ~1.0)