<h1 id="depth-pose-generator-for-krea-2-by-astroburner">🩶 Depth Pose Generator for Krea 2 by Astroburner</h1>Hey everyone!This time I wanted to create something completely different.Instead of training another character, style, outfit, or visual-effects LoRA, I experimented with a model that generates human pose depth maps directly from text prompts.I had seen a similar concept before and immediately wondered:What if I could describe a pose, generate the depth map, and reuse it as a control image for another generation?That idea became AstroDepthPoseV01 — and the first results are surprisingly strong. 🔥<hr /><h2 id="what-does-this-lora-do">🧠 What Does This LoRA Do?</h2>AstroDepthPoseV01 generates grayscale human depth maps based on natural-language pose descriptions.Instead of creating a finished character image, it produces a reusable spatial guide containing:<ul><li>Body posture and limb placement</li><li>Camera angle and perspective</li><li>Foreshortening</li><li>Body overlap</li><li>Front-to-back distance</li><li>Relative depth across the figure</li><li>Dynamic full-body poses</li></ul>The generated image can then be used with a compatible Depth ControlNet or another depth-guided workflow.The same depth pose can later become a realistic person, fantasy warrior, cyborg, alien, monster, anime character, superhero, or almost anything else your target model can generate.<hr /><h2 id="training-information">🏋️ Training Information</h2>The LoRA was trained with Krea 2 RAW and is intended primarily for generation with Krea 2 Turbo.The training dataset contained:<ul><li>281 manually selected depth-map images</li><li>58 different pose categories</li><li>Standing, sitting, kneeling, jumping, dancing, yoga, parkour, martial arts, climbing, balancing, boxing, gymnastics, surfing, and more</li><li>Portrait, landscape, and square compositions</li><li>Detailed individual captions describing pose geometry and spatial relationships</li></ul>Every training image contained one person and was manually reviewed before training.The dataset focuses on body geometry, pose, perspective, and relative depth rather than identity, clothing color, facial details, or photographic style.<hr /><h2 id="activation-tag">🏷️ Activation Tag</h2>Use:AstroDepthPoseV01A reliable prompt opening is:A human pose depth map, AstroDepthPoseV01,After that, describe the person, pose, body direction, limb placement, camera angle, and spatial relationships in normal English.<hr /><h2 id="prompting-tips">✍️ Prompting Tips</h2>The LoRA responds best when the prompt clearly describes:<ul><li>Full-body or partial-body framing</li><li>Standing, sitting, kneeling, lying, jumping, or balancing</li><li>Torso direction</li><li>Arm and leg placement</li><li>Camera position</li><li>Which body part is closest to the camera</li><li>Foreshortening and overlapping limbs</li><li>Furniture or environmental objects when required</li></ul>Detailed natural-language prompts currently produce the strongest results.<hr /><h2 id="example-dynamic-lunge">🩶 Example — Dynamic Lunge</h2>A human pose depth map, AstroDepthPoseV01, showing one adult person in a complete full-body view performing a deep forward lunge. The front leg bends strongly at the knee while the rear leg extends backward in a long line. The torso leans slightly forward and both arms reach diagonally upward above the shoulders. The camera views the figure from a low three-quarter angle, making the front thigh and nearest forearm appear larger and closer to the viewer. Continuous grayscale depth values represent the spatial structure, with strong separation between the forward limbs and the more distant rear leg.<hr /><h2 id="example-person-with-furniture">🪑 Example — Person with Furniture</h2>A human pose depth map, AstroDepthPoseV01, showing one adult person seated on the edge of a sofa in a complete full-body view. The torso leans slightly forward, one forearm rests across the thigh, and the other hand is raised near the face. Both legs angle forward, with one foot positioned closer to the viewer. The sofa occupies the midground and a low table appears farther into the scene. Continuous grayscale depth values define the seated body clearly against the furniture and surrounding room space.<hr /><h2 id="experimental-two-person-prompting">👥 Experimental Two-Person Prompting</h2>The training dataset contains only single-person images, but the LoRA can also be tested with two-person prompts.A human pose depth map, AstroDepthPoseV01, showing two adult people standing face to face in a complete full-body view. One person stands slightly closer to the camera and extends a hand forward, while the other person stands farther back and reaches toward the gesture. Their torsos angle toward each other. Continuous grayscale depth values describe both bodies separately, with the nearer figure appearing brighter and the more distant figure appearing darker.Two-person generation should currently be considered experimental.<hr /><h2 id="background-and-furniture-tests">🛋️ Background and Furniture Tests</h2>The LoRA can also generate simple environmental depth elements such as:<ul><li>Chairs</li><li>Sofas</li><li>Tables</li><li>Beds</li><li>Kitchen counters</li><li>Shelves</li><li>Stairs</li><li>Interior room structures</li></ul>Simple and clearly described environments usually work better than heavily overloaded scenes.<hr /><h2 id="suggested-lora-strength">⚙️ Suggested LoRA Strength</h2>Try several strengths depending on your workflow:<ul><li>0.6 — lighter influence</li><li>0.8 — balanced control</li><li>1.0 — strong depth-map appearance</li><li>1.2 — experimental stronger influence</li></ul>Results may vary depending on resolution, seed, guidance, sampler, and generation interface.<hr /><h2 id="experimental-first-version">⚠️ Experimental First Version</h2>This LoRA generates visually plausible relative depth maps. It does not produce scientifically measured or metrically accurate depth data.Complex overlaps, multiple people, unusual anatomy, and detailed environments may still create inconsistent depth relationships.However, the current results already appear highly useful as creative pose and composition guides.<hr /><h2 id="why-this-is-interesting">🚀 Why This Is Interesting</h2>Most depth workflows begin with an existing image.This LoRA reverses that process:Text prompt → generated depth map → controlled final imageThat makes it possible to invent new poses and spatial compositions without first searching for a matching reference photograph.<hr /><h2 id="recommended-resolution">🖼️ Recommended Resolution</h2>For the most stable results, generate at approximately 1 megapixel.Generation at up to 2 megapixels also works well when using a 1:1 aspect ratio. This configuration has been tested successfully.Going beyond 2 megapixels can introduce visible ghosting, duplicated contours, and unstable depth structures. For this reason, higher resolutions are currently not recommended.Recommended settings:<ul><li>Best overall stability: approximately 1 megapixel</li><li>Tested maximum: 2 megapixels at 1:1</li><li>Recommended aspect ratio at 2 megapixels: 1:1</li><li>Above 2 megapixels: increased risk of ghosting and duplicated depth shapes</li></ul>Created by Astroburner.

Depth Pose Generator for Krea 2 by Astroburner

Version 1.0Trained on 281 manually selected depth-map images across 58 human pose categories. The LoRA was trained with Krea 2 RAW for 3,000 steps, with 15 checkpoints, and is intended primarily for generation with Krea 2 Turbo.

Files with SHA256: 6b722f81eba5cbc5d3dffb75630bda185dc980d59cf53dfef668fd72e570ba42