Qwen-Edit-Krea2-Turbo-Workflow

Qwen-Edit → Krea 2 Turbo — 2-Stage Pose-to-Photoreal Workflow (ComfyUI)

A simple two-stage ComfyUI workflow that turns a pose reference into a photoreal result — the pose stays exact, the look is fully controllable. Same pose skeleton, any subject and scene.

Qwen owns the pose, Krea owns the look.

Gallery

All from the same pose skeletons — different subjects and scenes, one seed through the pipeline.

How it works

Stage 1 — Qwen-Image-Edit 2511 (structure & pose). A reference is reduced to its first frame, run through DWPose to get an OpenPose skeleton, and Qwen-Edit generates the subject strictly on that pose. Pixel-accurate posing, full freedom over character and scene from the prompt.

Stage 2 — Krea 2 Turbo (img2img, denoise ~0.42). The Qwen latent is refined by Krea 2 Turbo, which adds photoreal surface detail, lighting and polish on top of Qwen's structure. Each stage has its own separate prompt — Qwen gets a short pose-focused instruction, Krea gets a rich visual description.

Sample settings

• Stage 1 — Qwen-Edit: CFG 4, 20 steps, 1088×1920
• Stage 2 — Krea 2 Turbo: CFG 1, 32 steps, 1088×1920

Hardware / VRAM

Tuned for a big card — RTX PRO 6000, 96 GB — so both models stay resident at once for speed.

Modifies easily for lower VRAM (slower, but it works):

• Add an "Unload Model" node for Qwen right after the Qwen stage, before Krea loads → one model in VRAM at a time, fits one-by-one on smaller cards.
• Combine with fp8 weights, sequential model loading, tiled VAE.
• Rule of thumb: if both models don't fit together, unload between stages and run them one at a time.

Qwen-Edit takes up to 3 input images

The Qwen-Edit stage supports up to three input images (Picture 1 / 2 / 3). Bypass the ones you don't need, or enable them and reference each in the prompt ("subject of picture 1 in the pose of picture 2"). Typical use: one = subject, one = pose. Add a third for a style or background reference.

Control is a swappable module (build it like Lego)

The pose input is not fixed to OpenPose. The control block is just preprocessor → encode → apply, so swap the preprocessor and feed any ordinary photo:

• Depth → Depth-Anything-V2 → 3D structure/volume
• Pose → DWPose / OpenPose → joint pose (this build)
• Canny / Lineart → outlines/composition
• Normal maps → surface orientation

Same three nodes downstream, only the preprocessor changes.

Example prompts

The actual prompt file used for these samples is included: example_prompts.txt.

• Per stage: a short pose-focused Qwen prompt + a rich visual Krea prompt.
• Several subjects across 2 poses — the two poses are two different OpenPose reference frames.
• Some subjects run with no LoRA; others use a character LoRA (put your trigger word in, or bypass it and describe the character in words).
• The header also has the memory/VRAM notes and the control-swapping cheatsheet.

Models

• Qwen-Image-Edit 2511 (+ Qwen3-VL text encoder + Qwen-Image VAE)
• Krea 2 Turbo
• DWPose (comfyui_controlnet_aux)

Files

• Qwen_Edit_Krea2Turbo.json — the workflow (drop into ComfyUI)
• example_prompts.txt — the prompt examples
• gallery/ — sample outputs + workflow screenshots

---

Built and shared by JahJedi.