We present SEED, an elaborate image tokenizer that empowers Large Language Models (LLMs) with the emergent ability to SEE and Draw at the same time. Research on image tokenizers has previously reached an impasse, as frameworks employing quantized visual tokens have lost prominence due to subpar performance and convergence in multimodal comprehension (compared to BLIP-2, etc.) or generation (compared to Stable Diffusion, etc.). Despite the limitations, we remain confident in its natural capacity to unify visual and textual representations, facilitating scalable multimodal training with LLM's original recipe. In this study, we identify two crucial principles for the architecture and training of SEED that effectively ease subsequent alignment with LLMs. (1) Image tokens should be independent of 2D physical patch positions and instead be produced with a 1D causal dependency, exhibiting intrinsic interdependence that aligns with the left-to-right autoregressive prediction mechanism in LLMs. (2) Image tokens should capture high-level semantics consistent with the degree of semantic abstraction in words, and be optimized for both discriminativeness and reconstruction during the tokenizer training phase. As a result, the off-the-shelf LLM is able to perform both image-to-text and text-to-image generation by incorporating our SEED through efficient LoRA tuning. Comprehensive multimodal pretraining and instruction tuning, which may yield improved results, are reserved for future investigation. This version of SEED was trained in 5.7 days using only 64 V100 GPUs and 5M publicly available image-text pairs. Our preliminary study emphasizes the great potential of discrete visual tokens in versatile multimodal LLMs and the importance of proper image tokenizers in broader research.
While recent advancements in vision-language models have revolutionized multi-modal understanding, it remains unclear whether they possess the capabilities of comprehending the generated images. Compared to real data, synthetic images exhibit a higher degree of diversity in both content and style, for which there are significant difficulties for the models to fully apprehend. To this end, we present a large-scale dataset, JourneyDB, for multi-modal visual understanding in generative images. Our curated dataset covers 4 million diverse and high-quality generated images paired with the text prompts used to produce them. We further design 4 benchmarks to quantify the performance of generated image understanding in terms of both content and style interpretation. These benchmarks include prompt inversion, style retrieval, image captioning and visual question answering. Lastly, we assess the performance of current state-of-the-art multi-modal models when applied to JourneyDB, and provide an in-depth analysis of their strengths and limitations in generated content understanding. We hope the proposed dataset and benchmarks will facilitate the research in the field of generative content understanding. The dataset will be available on https://journeydb.github.io.
Text-guided image generation has witnessed unprecedented progress due to the development of diffusion models. Beyond text and image, sound is a vital element within the sphere of human perception, offering vivid representations and naturally coinciding with corresponding scenes. Taking advantage of sound therefore presents a promising avenue for exploration within image generation research. However, the relationship between audio and image supervision remains significantly underdeveloped, and the scarcity of related, high-quality datasets brings further obstacles. In this paper, we propose a unified framework 'Align, Adapt, and Inject' (AAI) for sound-guided image generation, editing, and stylization. In particular, our method adapts input sound into a sound token, like an ordinary word, which can plug and play with existing powerful diffusion-based Text-to-Image (T2I) models. Specifically, we first train a multi-modal encoder to align audio representation with the pre-trained textual manifold and visual manifold, respectively. Then, we propose the audio adapter to adapt audio representation into an audio token enriched with specific semantics, which can be injected into a frozen T2I model flexibly. In this way, we are able to extract the dynamic information of varied sounds, while utilizing the formidable capability of existing T2I models to facilitate sound-guided image generation, editing, and stylization in a convenient and cost-effective manner. The experiment results confirm that our proposed AAI outperforms other text and sound-guided state-of-the-art methods. And our aligned multi-modal encoder is also competitive with other approaches in the audio-visual retrieval and audio-text retrieval tasks.
Video Question Answering (VideoQA) has been significantly advanced from the scaling of recent Large Language Models (LLMs). The key idea is to convert the visual information into the language feature space so that the capacity of LLMs can be fully exploited. Existing VideoQA methods typically take two paradigms: (1) learning cross-modal alignment, and (2) using an off-the-shelf captioning model to describe the visual data. However, the first design needs costly training on many extra multi-modal data, whilst the second is further limited by limited domain generalization. To address these limitations, a simple yet effective Retrieving-to-Answer (R2A) framework is proposed.Given an input video, R2A first retrieves a set of semantically similar texts from a generic text corpus using a pre-trained multi-modal model (e.g., CLIP). With both the question and the retrieved texts, a LLM (e.g., DeBERTa) can be directly used to yield a desired answer. Without the need for cross-modal fine-tuning, R2A allows for all the key components (e.g., LLM, retrieval model, and text corpus) to plug-and-play. Extensive experiments on several VideoQA benchmarks show that despite with 1.3B parameters and no fine-tuning, our R2A can outperform the 61 times larger Flamingo-80B model even additionally trained on nearly 2.1B multi-modal data.
Recent progress on vision-language foundation models have brought significant advancement to building general-purpose robots. By using the pre-trained models to encode the scene and instructions as inputs for decision making, the instruction-conditioned policy can generalize across different objects and tasks. While this is encouraging, the policy still fails in most cases given an unseen task or environment. To adapt the policy to unseen tasks and environments, we explore a new paradigm on leveraging the pre-trained foundation models with Self-PLAY and Self-Describe (SPLAYD). When deploying the trained policy to a new task or a new environment, we first let the policy self-play with randomly generated instructions to record the demonstrations. While the execution could be wrong, we can use the pre-trained foundation models to accurately self-describe (i.e., re-label or classify) the demonstrations. This automatically provides new pairs of demonstration-instruction data for policy fine-tuning. We evaluate our method on a broad range of experiments with the focus on generalization on unseen objects, unseen tasks, unseen environments, and sim-to-real transfer. We show SPLAYD improves baselines by a large margin in all cases. Our project page is available at https://geyuying.github.io/SPLAYD/
Pre-training on large-scale video data has become a common recipe for learning transferable spatiotemporal representations in recent years. Despite some progress, existing methods are mostly limited to highly curated datasets (e.g., K400) and exhibit unsatisfactory out-of-the-box representations. We argue that it is due to the fact that they only capture pixel-level knowledge rather than spatiotemporal commonsense, which is far away from cognition-level video understanding. Inspired by the great success of image-text pre-training (e.g., CLIP), we take the first step to exploit language semantics to boost transferable spatiotemporal representation learning. We introduce a new pretext task, Turning to Video for Transcript Sorting (TVTS), which sorts shuffled ASR scripts by attending to learned video representations. We do not rely on descriptive captions and learn purely from video, i.e., leveraging the natural transcribed speech knowledge to provide noisy but useful semantics over time. Furthermore, rather than the simple concept learning in vision-caption contrast, we encourage cognition-level temporal commonsense reasoning via narrative reorganization. The advantages enable our model to contextualize what is happening like human beings and seamlessly apply to large-scale uncurated video data in the real world. Note that our method differs from ones designed for video-text alignment (e.g., Frozen) and multimodal representation learning (e.g., Merlot). Our method demonstrates strong out-of-the-box spatiotemporal representations on diverse video benchmarks, e.g., +13.6% gains over VideoMAE on SSV2 via linear probing.
Dominant pre-training work for video-text retrieval mainly adopt the "dual-encoder" architectures to enable efficient retrieval, where two separate encoders are used to contrast global video and text representations, but ignore detailed local semantics. The recent success of image BERT pre-training with masked visual modeling that promotes the learning of local visual context, motivates a possible solution to address the above limitation. In this work, we for the first time investigate masked visual modeling in video-text pre-training with the "dual-encoder" architecture. We perform Masked visual modeling with Injected LanguagE Semantics (MILES) by employing an extra snapshot video encoder as an evolving "tokenizer" to produce reconstruction targets for masked video patch prediction. Given the corrupted video, the video encoder is trained to recover text-aligned features of the masked patches via reasoning with the visible regions along the spatial and temporal dimensions, which enhances the discriminativeness of local visual features and the fine-grained cross-modality alignment. Our method outperforms state-of-the-art methods for text-to-video retrieval on four datasets with both zero-shot and fine-tune evaluation protocols. Our approach also surpasses the baseline models significantly on zero-shot action recognition, which can be cast as video-to-text retrieval.
Mainstream Video-Language Pre-training models \cite{actbert,clipbert,violet} consist of three parts, a video encoder, a text encoder, and a video-text fusion Transformer. They pursue better performance via utilizing heavier unimodal encoders or multimodal fusion Transformers, resulting in increased parameters with lower efficiency in downstream tasks. In this work, we for the first time introduce an end-to-end video-language model, namely \textit{all-in-one Transformer}, that embeds raw video and textual signals into joint representations using a unified backbone architecture. We argue that the unique temporal information of video data turns out to be a key barrier hindering the design of a modality-agnostic Transformer. To overcome the challenge, we introduce a novel and effective token rolling operation to encode temporal representations from video clips in a non-parametric manner. The careful design enables the representation learning of both video-text multimodal inputs and unimodal inputs using a unified backbone model. Our pre-trained all-in-one Transformer is transferred to various downstream video-text tasks after fine-tuning, including text-video retrieval, video-question answering, multiple choice and visual commonsense reasoning. State-of-the-art performances with the minimal model FLOPs on nine datasets demonstrate the superiority of our method compared to the competitive counterparts. The code and pretrained model have been released in https://github.com/showlab/all-in-one.
Dancing video retargeting aims to synthesize a video that transfers the dance movements from a source video to a target person. Previous work need collect a several-minute-long video of a target person with thousands of frames to train a personalized model. However, the trained model can only generate videos of the same person. To address the limitations, recent work tackled few-shot dancing video retargeting, which learns to synthesize videos of unseen persons by leveraging a few frames of them. In practice, given a few frames of a person, these work simply regarded them as a batch of individual images without temporal correlations, thus generating temporally incoherent dancing videos of low visual quality. In this work, we model a few frames of a person as a series of dancing moves, where each move contains two consecutive frames, to extract the appearance patterns and the temporal dynamics of this person. We propose MetaDance, which utilizes temporal-aware meta-learning to optimize the initialization of a model through the synthesis of dancing moves, such that the meta-trained model can be efficiently tuned towards enhanced visual quality and strengthened temporal stability for unseen persons with a few frames. Extensive evaluations show large superiority of our method.
Pre-training a model to learn transferable video-text representation for retrieval has attracted a lot of attention in recent years. Previous dominant works mainly adopt two separate encoders for efficient retrieval, but ignore local associations between videos and texts. Another line of research uses a joint encoder to interact video with texts, but results in low efficiency since each text-video pair needs to be fed into the model. In this work, we enable fine-grained video-text interactions while maintaining high efficiency for retrieval via a novel pretext task, dubbed as Multiple Choice Questions (MCQ), where a parametric module BridgeFormer is trained to answer the "questions" constructed by the text features via resorting to the video features. Specifically, we exploit the rich semantics of text (i.e., nouns and verbs) to build questions, with which the video encoder can be trained to capture more regional content and temporal dynamics. In the form of questions and answers, the semantic associations between local video-text features can be properly established. BridgeFormer is able to be removed for downstream retrieval, rendering an efficient and flexible model with only two encoders. Our method outperforms state-of-the-art methods on the popular text-to-video retrieval task in five datasets with different experimental setups (i.e., zero-shot and fine-tune), including HowTo100M (one million videos). We further conduct zero-shot action recognition, which can be cast as video-to-text retrieval, and our approach also significantly surpasses its counterparts. As an additional benefit, our method achieves competitive results with much shorter pre-training videos on single-modality downstream tasks, e.g., action recognition with linear evaluation.