Add JoyAI-Image Edit Plus pipeline and model

[tangyanf]

Model/Pipeline/Scheduler description

Model description

JoyAI-Image is a unified multimodal foundation model for image understanding, text-to-image generation, and instruction-guided image editing from JD.com. It combines an 8B
Multimodal Large Language Model (MLLM) with a 16B Multimodal Diffusion Transformer (MMDiT).

JoyAI-Image Edit Plus is the multi-image instruction-guided editing variant. Unlike the single-image JoyAI-Image Edit (added in #XXXX), Edit Plus accepts 1–6 reference images and a
text instruction to generate a new image that combines elements from the references.

Key architectural differences from Edit:

• Patchified 6D latent representation: Input images and target noise are independently VAE-encoded and patchified into [B, max_patches, C, pt, ph, pw] format with a target_mask to
  distinguish target noise from reference patches.
• Variable reference images: Supports 1–6 reference images per sample via dynamic shape_list.
• Batched RoPE: Per-component rotary position embeddings with temporal offsets for each reference image and the target.
• Norm-guided CFG: Classifier-free guidance with norm rescaling in a single forward pass.

Open source status

• The model implementation is available.
• The model weights are available (Only relevant if addition is not a scheduler).

Provide useful links for the implementation

• GitHub: https://github.com/jd-opensource/JoyAI-Image
• Weights (Edit): https://huggingface.co/jdopensource/JoyAI-Image-Edit
• Weights (Edit Plus): https://huggingface.co/jdopensource/JoyAI-Image-Edit-Plus-Diffusers