Identity Master – Experimental Playground Workflow

Concept and Objective

This ComfyUI workflow explores the idea of maintaining a consistent visual identity across different image generations while allowing poses, camera perspectives, compositions, and environments to change freely.

At its core is an Identity Master, which serves as the primary visual reference for a person's appearance. AI-powered image analysis is used to extract defining characteristics such as facial features, hair, skin, and body proportions. Independently, the workflow analyzes up to three pose reference images, attempting to reconstruct their body positions, spatial geometry, camera perspectives, and compositions as accurately as possible.

The extracted information is then automatically combined into individual prompts optimized for Krea2. Additional experimental techniques include DWPose/OpenPose, AI-assisted pose geometry validation, and optional clothing overrides.

The main objective is to separate a person's visual identity from their pose and environment as reliably as possible. This should make it possible to generate different image variations without having the model reinterpret the person's appearance with every generation.

Experimental Nature

This workflow is explicitly designed as an experimental playground rather than a finished, production-ready solution. Its purpose is to explore different approaches, models, prompting strategies, and processing stages, combine them in various ways, and compare their results directly.

The focus is on experimentation, identifying technical limitations, and continuously improving identity consistency, pose accuracy, and overall image quality.

Rather than being a finalized system, the workflow serves as a flexible testing environment where new ideas can be integrated quickly, existing approaches can be challenged, and different techniques can be evaluated through hands-on experimentation.

Advanced Approach V2.0: In this version, I’ve integrated a DWPose Estimator node and a Keypoint Interpreter into the pose validation pipeline. The Keypoint Interpreter extracts and interprets skeletal keypoint data to provide additional geometric context, while the Validator cross-checks this information against the original reference image to identify and correct pose inaccuracies.

The final merge stage combines the identity features from the Identity Master, the validated pose description, and the optional Clothing Lock into a single prompt.

Changelog: Generated Text Playground v2.0 → v3.0

Focus: More accurate pose reconstruction, dedicated camera and composition analysis, and improved AI-powered validation.

What's New and What's Changed

• Sapiens2 Integration: Replaced the previous DWPose detection with Sapiens2, including RT-DETR person detection and skeleton visualization.

• Camera & Composition Extractor: Added a dedicated analysis stage for camera perspective, framing, and image composition.

• Camera & Composition Validator: Introduced an additional AI-powered validation step to review and correct extracted camera and composition details.

• Reworked Pose Extraction: Expanded the system prompts to improve the reconstruction of body posture, limb positioning, overlaps, and spatial relationships.

• Improved Pose Validation: Added stricter checks for anatomical accuracy, image-space coordinates, and potentially unreliable skeleton data.

• Enhanced Prompt Merger: Reworked how identity, validated pose information, and camera and composition details are combined.

• Optimized AI Parameters: Standardized generation settings with lower temperature values and reduced maximum output lengths.

• Updated Reference Processing: Increased the target resolution for pose references from 512 × 751 to 768 × 1126 pixels.

Summary

Version 3.0 takes the experimental workflow a step further by introducing Sapiens2-based pose detection and a separate, multi-stage camera and composition analysis pipeline. Both pose validation and prompt merging have been significantly reworked.

The goal is to achieve more reliable pose reconstruction while preserving the original identity and visual composition.

ComfyUI-Sapiens2 Git: https://github.com/kijai/ComfyUI-Sapiens2
Sapiens2 Model Page (sapiens2_1b_pose.safetensors): https://huggingface.co/facebook/sapiens2-pose-1b/blob/main/README.md
RT-DETRv4 diffusion_models - (rt_detr_v4-x-hgnet_fp16.safetensors): https://huggingface.co/Comfy-Org/SDPose/tree/main/diffusion_models