[feat] Introduce cross-modality and multi-modality support; modularize CrossEncoder class

[tomaarsen]

Hello!

Pull Request overview

• Introduce cross-modality and multi-modality support via refactors of SentenceTransformer, Router, and Transformer
• Modularize the CrossEncoder class, initially by subclassing SentenceTransformer, but long term I want to subclass a new superclass

Details

This pull request is very much a work-in-progress, although it is already functional. In short:

1. Transformer now works with an AutoProcessor and handles inputs through that. This accepts multiple modalities
2. SentenceTransformer, Transformer and Router check the modality of inputs, only one modality is allowed per inference
3. Router has been adapted to allow for modality-based routing
4. There is a strict distinction between a model with modalities ["text", "image"] and [("text", "image")]. The former is cross-modal, i.e. you can pass either text or images, and you can then compare the embeddings across the modalities. The latter is multi-modal, i.e. you can pass text AND images at the same time, and this joint input results in one embedding output. The "one input in, one embedding out" is a core feature.
5. Multimodal models can be called with lists of dictionaries using modalities as keys, e.g. model.encode([{"text": "This is my <image>", "image": "cat.jpg"}, ...]).
6. Transformer is designed to be somewhat flexible moving forward. Model authors can specify which modalities are supported, which methods on the AutoModel need to be called, and which output keys need to be used from the outputs from those methods. The goal is to have strong defaulting as well.
7. model.modalities gives a list of supported modalities. E.g. SentenceTransformer("laion/clap-htsat-unfused").modalities is ['text', 'audio']
8. The modalities are text, image, audio, video, and combination of the previous, and message. The latter uses processor.apply_chat_template.

Here's two cross-modal models that I trained:

• https://huggingface.co/tomaarsen/clap-htsat-fused-librispeech-cont-4-epochs-128bs
• https://huggingface.co/tomaarsen/clip-vit-L14-coco

This is an incomplete list of models that you can simply initialize with SentenceTransformer(model_name):

• text
  • all-MiniLM-L6-v2
  • google/embeddinggemma-300m
  • Qwen/Qwen3-Embedding-0.6B
  • google/gemma-3-1b-pt
• image
  • google/vit-base-patch16-224-in21k
  • facebook/deit-base-distilled-patch16-224
  • facebook/dinov2-with-registers-small
  • DeepGlint-AI/mlcd-vit-bigG-patch14-336
  • microsoft/resnet-18
  • timm/mobilenetv4_conv_medium.e500_r256_in1k
  • timm/convnext_base.clip_laion2b
  • microsoft/beit-base-patch16-224
  • google/bit-50
  • microsoft/conditional-detr-resnet-50
• audio
  • facebook/wav2vec2-large-960h-lv60-self
  • nari-labs/Dia-1.6B-0626
  • facebook/hubert-large-ls960-ft
• cross text+image
  • kakaobrain/align-base
  • apple/aimv2-large-patch14-224-lit
  • openai/clip-vit-base-patch32
  • google/siglip-base-patch16-224
• cross text+audio
  • laion/clap-htsat-unfused
• multi text+image
  • google/paligemma-3b-mix-448
  • ds4sd/SmolDocling-256M-preview
  • ibm-granite/granite-docling-258M
  • ibm-granite/granite-vision-3.3-2b
• multi text+image as message
  • deepseek-community/deepseek-vl-1.3b-chat

For more complex setups, you can use a Router, e.g. when one transformer model doesn't have the modalities that you're after

from PIL import Image
from sentence_transformers import SentenceTransformer
from sentence_transformers.models import Dense, Pooling, Router, Transformer

# Create separate encoders for different modalities
text_encoder = Transformer("sentence-transformers/all-MiniLM-L6-v2")
# Project to 768 dims to match image encoder
text_dense = Dense(text_encoder.get_word_embedding_dimension(), 768, module_input_name="token_embeddings")
image_encoder = Transformer(
    "ModernVBERT/modernvbert",
    model_args={"trust_remote_code": True},
    tokenizer_args={"trust_remote_code": True},
    config_args={"trust_remote_code": True},
)
pooling = Pooling(text_encoder.get_word_embedding_dimension())

# Route based on modality
router = Router(
    sub_modules={
        "text": [text_encoder, text_dense],
        "image": [image_encoder],
    },
    route_mappings={
        (None, "text"): "text",  # Any task with text goes to text encoder
        (None, ("text", "image")): "image",  # Any task with text-image together goes to image encoder
    },
)

model = SentenceTransformer(modules=[router, pooling])

# Modality is automatically inferred
text_embedding = model.encode("A photo of a cat")
multimodal_embedding = model.encode({"text": "A photo of a <image>", "image": Image.open("cat.jpg")})

similarity = model.similarity(text_embedding, multimodal_embedding)

For the text modality, this example uses all-MiniLM-L6-v2 with a linear layer that projects the token embeddings to 768-dimensional before pooling. For the multimodal text+image, this uses ModernVBERT/modernvbert, a model that supports both text AND image inputs simultaneously for one output embedding. This model could be trained to perform multimodal retrieval or similar tasks.

There are currently a lot of tiny breaking changes that I want to iron out. If I can't get rid of all of them, then sadly this refactor will have to wait until a v6.0 release, which I would normally only do alongside the introduction of a new archetype, Late Interaction in this case. Tons of TODOs also remain.

cc @NohTow - this should in theory allow you to work on multi-modal/cross-modal Late Interaction. One big annoyance for you for now will likely be that many architectures like CLIP/CLAP default to using get_text_features/get_..._features from the transformers model, and these methods all output pooled embeddings rather than token embeddings. This is okay-ish for Sentence Transformers, but a big problem for LI models.

• Tom Aarsen