🤗 Model   |    💬 Demo: Chat with Dimple   |   📑 Paper   |    ✨ Code  

💧 Dimple

Dimple is the first Discrete Diffusion Multimodal Large Language Model (DMLLM) that leverages a hybrid training paradigm combining autoregressive and diffusion-based instruction tuning. The model architecture is similar to Qwen and LLaVA, while introducing an autoregressive-then-diffusion training strategy:

• Stage 1: Autoregressive fine-tuning for alignment and initial instruction tuning.
• Stage 2: Diffusion-based fine-tuning for enhanced instruction-following capabilities.

Trained on the same dataset as LLaVA-NEXT, Dimple-7B surpasses LLaVA-NEXT-7B by 3.9%, demonstrating that diffusion-based multimodal large language models can match its autoregressive counterparts under similar training budget.

---

🔍 Highlights

• Hybrid Training: Combines autoregressive and diffusion training.
• Diffusion Decoding: Supports confident decoding, random decoding, maskgit-style decoding, and entropy-based decoding.
• Controllable Generation: Enables fine-grained control over format, structure, and length via structure priors.
• Autoregressive-like Prefilling: Enhances inference speed using prefilling techniques.

---

📊 Evaluation Results

---

🛠️ Environment

Make sure your environment includes the following versions:

transformers==4.46.2
torch==2.5.1
accelerate==1.6.0

---

🚀 Inference Example

import torch
from transformers import AutoProcessor, AutoModel
import json, requests
from PIL import Image

model_name = "rp-yu/Dimple-7B"
processor = AutoProcessor.from_pretrained(
    model_name,
    trust_remote_code=True
)
model = AutoModel.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
)

image_url = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"
messages = [
    [{"role": "user", "content": [
        {"type": "image", "image": image_url},
        {"type": "text", "text": "Describe this image."}
    ]}],
]
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True, add_vision_id=False
)
images = [
    Image.open(requests.get(image_url, stream=True).raw).convert("RGB")
]

inputs = processor(
    text=text,
    images=images,
    videos=None,
    padding="longest",
    return_tensors="pt",
)

input_ids = inputs.pop("input_ids")
output = model.diffusion_generate(
    input_ids,
    max_new_tokens=64,
    output_history=True,
    return_dict_in_generate=True,
    steps=64,
    temperature=0.2,
    top_p=0.95,
    alg="origin",
    use_cache=True,
    alg_p_threshold=0.95,
    use_original_confidence=True,
    decoding_pipeline="dim",
    **inputs
)

generations = [
    processor.tokenizer.decode(g[len(p):].cpu().tolist())
    for p, g in zip(input_ids, output.sequences)
]

for j in range(len(messages)):
    print("output:", j, generations[j].split(processor.tokenizer.eos_token)[0])

# output: 0 In the image, a woman wearing a shirt with a plaid and a dog are sitting together on a beach. The sun appears to be setting in the background, creating a warm and serene atmosphere.

---

🏆 Evaluation

We evaluate Dimple using lmms-eval. The evaluation scripts are provided in the lmms-eval folder, and the evaluation commands for Dimple can be found at lmms-eval/examples/models/dimple.sh.

To run the evaluation, follow these steps:

1. Install Dimple dependencies Make sure all required packages for Dimple are installed. Refer to Dimple’s environment setup instructions for details.

2. Install lmms-eval dependencies Next, install the necessary dependencies for lmms-eval. These are listed in the lmms-eval repository.

3. Set your OPENAI_API_KEY Some tasks require OpenAI API access. To do this, edit the file lmms-eval/examples/models/dimple.sh and replace OPENAI_API_KEY=MY_OPENAI_API_KEY with your actual API key.

Once all dependencies are installed and your API key is set, you can run the evaluation script directly:

sh lmms-eval/examples/models/dimple.sh

This will execute the evaluation pipeline for Dimple using the default configuration.

---

💪 Training

The training code is provided in training-code. Before using the code, you should also prepare your training data and set up the base model.

• Training Data: Please follow the instruction in LLaVA-Next to prepare the data for alignment and instruction tuning.
• Base Model: The LLM is initialized with Dream-org/Dream-v0-Instruct-7B. The vision encoder is initialized with Qwen/Qwen2-VL-7B-Instruct. You may use the code in training-code/init.py to set up the base model.

---

💬 Local Demo

We can no longer afford to host the Hugging Face online demo. However, the code to run our demo is available in the demo directory, allowing you to run it locally and experience Dimple.

---

📚 Citation

@inproceedings{dimple,
    title={Dimple: Discrete Diffusion Multimodal Large Language Model with Parallel Decoding}, 
    author={Runpeng Yu and Xinyin Ma and Xinchao Wang},
    booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    year = "2026",
    publisher = "Association for Computational Linguistics",
}