<h2 id="design-philosophy">Design Philosophy</h2><p>This workflow is intended as a reference implementation. It prioritizes clarity, modularity and low VRAM compatibility over automation, making it easy to understand, modify and extend.</p><h2 id="purpose">Purpose</h2><p>• Blend the identity of a Character Reference into an Environment Reference.</p><p>• Optimized for Flux2Klein GGUF.</p><p>• Low VRAM friendly.</p><h2 id="workflow-steps">Workflow Steps</h2><ol><li><p>Load Character Reference</p></li><li><p>Load Environment Reference</p></li><li><p>Adjust Positive Prompt</p></li><li><p>Adjust Character Mask Settings (Enable / Disable relevant switches)</p></li><li><p>Adjust Environment Mask settings (only if needed)</p></li><li><p>Queue Prompt</p></li></ol><h2 id="notes">Notes</h2><p>This workflow blends two reference images using Flux2Klein's reference conditioning. In general, the first reference has a stronger influence on identity while the second reference contributes composition and scene context, although the final blend depends on the compatibility of both references and the model's learned priors.</p><p>• Character identity comes primarily from Character Reference.</p><p>• Scene composition comes primarily from Environment Reference.</p><p>• Prompt refines the final result.</p><p>• Use high-quality references for best output.</p><h2 id="reference-selection">Reference Selection</h2><p>• This workflow is designed for character replacement by blending two reference images.</p><p>• Character Reference: The identity, facial features and appearance you want to preserve.</p><p>• Environment Reference: The composition, pose, camera angle, lighting and scene you want to inherit.</p><p>Best Results</p><p>• Use references with similar framing and aspect ratios.</p><p>• Similar poses and camera angles generally produce more natural replacements.</p><p>• Large differences in framing (Landscape vs Portrait), pose or composition may produce unexpected results.</p><h2 id="reference-images">Reference Images</h2><p>• Reference images are automatically normalized to a maximum dimension of 1280 px.</p><p>• This reduces encoding time and VRAM usage while preserving visual fidelity.</p><p>• Increase the maximum resolution for higher reference fidelity if additional GPU memory is available.</p><h2 id="mask-configuration">Mask Configuration</h2><p>• The default masking pipeline is optimized for realistic and near-realistic images.</p><p>• Stylized artwork (Anime, Manga, Cartoons, etc.) may require an alternative segmentation model or manual mask refinement.</p><p>• Depending on the reference image, you may need to adjust the face, hair or confidence settings for improved segmentation.</p><p>• Images with complex hairstyles, accessories or occlusions may require additional tuning.</p><p>• This workflow provides a validated baseline. Feel free to adapt the masking pipeline for your own datasets and use cases.</p>

Files with SHA256: f0030ece766975a34d4fe637f05a47d8e8fc8bf8e445a0670920c380ced282f4