This Forge extension replaces SDXL/Mugen prompt conditioning with Jina CLIP v2 plus either a V2 or redesigned V3 Jina-to-SDXL adapter.

This extension adds support for jina-clip-v2 adapter as a text encoder for Mugen.

• jina-clip-v2
• Adapter+UNet

Place or clone this directory under Forge Neo's extensions/ directory, then restart Forge.

• Put a local Jina CLIP v2 model under models/llm/, models/LLM/, models/Jina/, or models/text_encoder/. A Hugging Face model ID such as jinaai/jina-clip-v2 can also be entered in the UI.
• Put adapter .safetensors files under models/llm_adapter/, models/llm_adapters/, models/LLM_Adapters/, or models/Jina_Adapters/.

Open Jina CLIP v2 SDXL Adapter (V2/V3) in txt2img or img2img, enable it, and select the model and adapter.

The default auto architecture selection examines the checkpoint:

• redesigned V3 checkpoints contain layer_fusion.* and mean_pooled_projection.* tensors;
• V2 checkpoints contain the existing seq_projection.* and attention_blocks.* sequence path.

V3 captures Jina text layers h8 and h16 plus the final normalized h24 state, performs token-wise layer fusion, and uses the checkpoint's V3 pooled projection. A V3 checkpoint must be complete and shape-compatible; partial or mislabeled V3 checkpoints are rejected instead of being loaded silently.

The V2 max-sequence, attention-pooling, and positional-embedding controls apply only to V2. Convert legacy fused-MHA adapter converts older in_proj_weight / in_proj_bias checkpoints to the current explicit q/k/v layout. Current V2 and V3 checkpoints do not need it.

The optional cross-attention padding mask uses Forge's existing extension patch and works with either adapter version. Prompt weighting, @ artist formatting, CLIP-less Mugen checkpoint loading, and the Forge conditioning override remain compatible with V2.

• The Jina model must expose its 24-layer text transformer; V3 specifically requires h8, h16, and h24.
• The tokenizer is loaded with fix_mistral_regex=False, matching the Jina XLM-R token IDs used during V2/V3 adapter training.
• Loading uses trust_remote_code=True, as required by Jina CLIP v2.
• Local files only requires all Jina model and trusted-code dependencies to already exist locally.
• Adapter conditioning is emitted as SDXL-compatible 2048-wide cross-attention plus a 1280-wide pooled vector. When the padding mask is enabled, its extra channel is removed by the Forge attention patch before the U-Net projections.