Fix blocky/checkerboard artifacts: pass kv_cache=True so reference (pose) tokens use isolated attention as this LoRA was trained (matches ComfyUI Krea2OstrisEditModelPatch kv_cache=True)

Fix blocky/checkerboard artifacts: deterministic VAE reference-latent encode (posterior mode) matching ComfyUI, instead of stochastic latent_dist.sample()

Fix pose fidelity/quality: replicate ComfyUI FluxKontextImageScale (nearest ~1MP bucket + center-crop + Lanczos) for pose sizing so ref-latent grid == output grid at native ~1MP resolution

Match ComfyUI workflow: guidance_scale=0 (cfg=1/no-CFG on Turbo), single LoRA scaling, aspect-preserving <=1MP /16 sizing

Move theme/css to launch() per Gradio 6

Match output size to pose for stronger control + add interactive pose-editing (Pose Image / Regular Image tabs)

Upload app.py with huggingface_hub

Upload folder using huggingface_hub