Sub-issue of #88. Wire the full pipeline: identity (winning method) + style (IPAdapter) + ControlNet pose (OpenPose/Depth/Canny) + scene prompt that does NOT re-describe facial traits. Run the full test matrix (weight sweep, pose/action change, text-editability) and the 1–5 scoring (identity / style / prompt-compliance / artifact-overfitting) with weight+base+steps+seed logged per run. Baseline to beat = Robert's Qwen-Image-Edit cafe render (waxy, weak likeness). Success = a local, free, on-model character image beating the paid Nano/GPT-image baseline. Then feed the winning workflow to Palette #219 as a provider. Optional follow-up: stack existing FLUX LoRAs on the winner.
Sub-issue of #88. Wire the full pipeline: identity (winning method) + style (IPAdapter) + ControlNet pose (OpenPose/Depth/Canny) + scene prompt that does NOT re-describe facial traits. Run the full test matrix (weight sweep, pose/action change, text-editability) and the 1–5 scoring (identity / style / prompt-compliance / artifact-overfitting) with weight+base+steps+seed logged per run. Baseline to beat = Robert's Qwen-Image-Edit cafe render (waxy, weak likeness). Success = a local, free, on-model character image beating the paid Nano/GPT-image baseline. Then feed the winning workflow to Palette #219 as a provider. Optional follow-up: stack existing FLUX LoRAs on the winner.