DWPose Structure Control for Krea 2

After experimenting with depth-map generation, I wanted to explore whether Krea 2 could also learn to generate reusable DWPose-style structure maps.

The result is Astro DWPose Structure Control for Krea 2, a specialized LoRA designed to generate human pose structures that can later be used as pose-control references in ControlNet workflows.

I am releasing this LoRA for free as a small thank-you to the community.

Activation Trigger

AstroDWPoseSimpleV01

Place the activation trigger naturally near the beginning of your prompt.

What It Can Generate

The LoRA works best for human DWPose-style structures, including:

• standing and walking poses

• seated, kneeling and lying poses

• dynamic movement and jumping poses

• unusual arm and leg positions

• difficult floor poses

• two-person interactions

• experimental multi-person compositions

• visible hand and finger keypoints

• simplified head-direction indicators

Complex poses and multiple people are possible, although results may vary depending on the prompt, seed and LoRA strength.

Recommended Resolution

For the best pose accuracy, generate at approximately 1 megapixel or below.

Good starting resolutions include:

• 1024 × 1024

• 896 × 1152

• 1152 × 896

• 768 × 1344

• 1344 × 768

Generating directly at significantly higher resolutions may introduce:

• white patches or blank areas

• malformed or duplicated joints

• weaker pose accuracy

• distorted hands and limbs

• inconsistent structure colors

The LoRA produced noticeably cleaner and more complex poses at approximately 1 MP than at 2 MP.

Recommended Krea Workflow

1. Generate the initial DWPose structure at approximately 1 MP or below.

2. Apply a 1.5× latent upscale.

3. Use a light denoise refinement during the latent upscale.

4. Adjust the denoise strength depending on how closely you want to preserve the original pose.

This workflow improves:

• fine finger and hand keypoints

• small joint connections

• line clarity

• structure stability

• color consistency

The latent upscale is especially useful because the initial lower-resolution generation preserves the pose more reliably, while the refinement pass restores smaller details.

Prompt Structure

A useful prompt format is:

AstroDWPoseSimpleV01, a simplified DWPose structure map on a black background, [number of figures], [framing], [precise pose description], clear body keypoints, visible hand keypoints, simplified head direction indicators, clean colored pose lines and joints

Describe each person separately when generating multiple figures.

Useful details include:

• which leg supports the body

• which knees are bent

• arm direction

• hand placement

• torso lean and rotation

• head direction

• contact points between people

• whether the figure is standing, sitting, kneeling or lying down

Example Prompt

AstroDWPoseSimpleV01, a simplified DWPose structure map on a black background, two complete human figures performing a dynamic partner dance turn, the first figure stands in a wide stable stance with one arm lifted overhead, the second figure rotates beneath the raised arm with the torso twisting and one leg crossing behind the other, their connected hands form a clear contact point above their heads, both free arms extend outward, both heads turn toward the movement, clearly separated body keypoints, visible hand keypoints, simplified head direction indicators, precise colored pose lines and joints

Known Limitations

This is an experimental control-image LoRA rather than a normal character or style LoRA.

Please keep the following limitations in mind:

• It is designed primarily for human poses.

• Animal pose structures are not reliably supported.

• Prompting animals may cause Krea to generate parts of the actual animal instead of a usable pose structure.

• Detailed facial landmark generation was not reliable enough to be advertised as a supported feature.

• Multiple overlapping figures can occasionally merge limbs or keypoints.

• Very high native resolutions can reduce pose accuracy.

• Some seeds interpret complex interactions more successfully than others.

For challenging poses, testing several seeds is recommended.

Intended Use

The generated structures can be used as starting points for:

• pose ControlNet workflows

• character generation

• composition planning

• testing difficult body positions

• creating pose references

• combining generated pose maps with other LoRAs and workflows

Generate the pose structure first, upscale it in latent space if needed, and then use the resulting image as your control reference.

I would love to see what kinds of poses and workflows you create with it. Feel free to share your results, successful prompts and unusual experiments in the comments.

Created and tested by Astroburner AI.