Vital developments in massive language fashions (LLMs) have impressed the event of multimodal massive language fashions (MLLMs). Early MLLM efforts, similar to LLaVA, MiniGPT-4, and InstructBLIP, exhibit notable multimodal understanding capabilities. To combine LLMs into multimodal domains, these research explored projecting options from a pre-trained modality-specific encoder, similar to CLIP, into the enter area of LLMs, enabling multimodal understanding and reasoning throughout the transformer spine. Though there are numerous design decisions for MLLMs, similar to imaginative and prescient encoders, characteristic alignment adapters, and datasets, the coaching for many of those fashions adheres to the autoregressive era paradigm, which has confirmed efficient for textual content era in LLMs. Regardless of their robust multimodal understanding capabilities, these fashions primarily deal with visible notion and lack the power to generate multimodal outputs past textual content.
Transformer fashions have demonstrated nice success in autoregressive modeling in pure language processing. Impressed by such progress, earlier research have instantly utilized the identical autoregressive modeling to be taught the dependency of picture pixels for picture and video era. For example, VideoPoet employs a decoder-only transformer structure to synthesize high-quality movies from multimodal inputs. Extra lately, LlamaGen has proven that a big language mannequin structure like Llama can autoregressively mannequin picture tokens, reaching first rate efficiency in class-conditional picture era.
On this article, we are going to talk about Present-O, a unified transformer that integrates multimodal understanding and era. In contrast to absolutely autoregressive fashions, Present-O unifies autoregressive and discrete diffusion modeling to adaptively deal with inputs and outputs of assorted and blended modalities. The unified mannequin flexibly helps a variety of vision-language duties, together with visible query answering, text-to-image era, text-guided inpainting/extrapolation, and mixed-modality era. Throughout varied benchmarks, Present-O demonstrates comparable or superior efficiency to current particular person fashions with an equal or bigger variety of parameters, highlighting its potential as a next-generation basis mannequin.
On this framework, the mannequin is tasked with predicting Gaussian noise added to the continual latent representations. In distinction, different fashions like D3PM, Masks-predict, ARDM, and MaskGIT use a discrete corruption course of as a substitute for Gaussian diffusion. Particularly, a picture is represented as a sequence of discrete tokens utilizing picture tokenizers, with every token related to a categorical label. The token-wise distribution is reworked right into a uniform distribution by a stochastic sampling course of. Throughout coaching, a portion of those tokens is randomly masked, and the mannequin is skilled to foretell the unique values of the masked tokens. On this work, Present-O adopts discrete diffusion modeling for visible era.
Over the previous few years, vital developments have emerged within the two key pillars of multimodal intelligence: understanding and era. For multimodal understanding, Multimodal Giant Language Fashions (MLLMs) like LLaVA have demonstrated distinctive capabilities in vision-language duties similar to visible question-answering (VQA). For visible era, denoising diffusion probabilistic fashions (DDPMs) have revolutionized conventional generative paradigms, reaching unprecedented efficiency in text-to-image/video era.
Given these achievements in particular person fields, it’s pure to discover the potential of connecting them. Current works have tried to assemble knowledgeable fashions from these two completely different domains to type a unified system that may deal with each multimodal understanding and era. Nevertheless, current makes an attempt usually contain separate fashions for understanding and era. For example, NExT-GPT employs a base language mannequin for multimodal understanding however requires an extra pre-trained diffusion mannequin for picture era. This raises the query: can one single transformer deal with each multimodal understanding and era?
Just lately, Chameleon demonstrated that that is potential. Particularly, Chameleon allows the fusion of various modalities to generate each textual content and picture tokens by autoregressive modeling. Whereas it is sensible to mannequin textual content tokens autoregressively, it’s much less clear whether or not modeling picture patches or pixels in the identical means is perfect. A key bottleneck of autoregressively predicting a picture is the big variety of sampling steps required, particularly when coping with greater decision photographs. Steady diffusion fashions have proven superior efficiency in visible era in comparison with autoregressive ones.
This leads us to discover whether or not a single transformer can combine each autoregressive and diffusion modeling. Present-O envisions a brand new paradigm the place textual content is represented as discrete tokens and modeled autoregressively, whereas steady picture pixels are modeled utilizing denoising diffusion. Nevertheless, integrating these two distinct methods right into a single community is non-trivial as a result of variations between discrete textual content tokens and steady picture representations. Moreover, diffusion fashions usually depend on two distinct fashions: a textual content encoder and a denoising community.
To handle this, Present-O introduces a novel unified mannequin able to dealing with each multimodal understanding and era duties utilizing blended autoregressive and diffusion modeling. Present-O is constructed upon a pre-trained LLM and leverages its autoregressive modeling capabilities for text-based reasoning. Impressed by different works, Present-O employs discrete denoising diffusion to mannequin picture tokens as a substitute of steady representations. Furthermore, Present-O inherently encodes textual content conditional data, eliminating the necessity for added textual content encoders. By using textual content and picture tokenizers, Present-O can course of numerous enter knowledge and duties, offering solutions autoregressively for vision-language duties and producing photographs utilizing discrete denoising diffusion.
Present-O demonstrates comparable, and in some instances higher, efficiency than particular person fashions with an equal or bigger variety of parameters throughout varied benchmarks. In contrast to autoregressive picture era, the Present-O framework requires about 20 instances fewer sampling steps, making it inherently sooner. Moreover, the Present-O framework helps downstream functions like text-guided inpainting and extrapolation with out requiring fine-tuning, as demonstrated within the following picture.
Present-O additionally has the potential for mixed-modality era, similar to interleaved video keyframe era with textual content descriptions, displaying promise for long-form video era. Moreover, the Present-O framework investigates the affect of discrete and steady picture representations on multimodal understanding, providing insights for future unified mannequin designs.
The next determine presents a comparability of mannequin traits between the Present-O framework and current strategies throughout varied domains. Present-O stands out as a unified mannequin that integrates superior methods for each multimodal understanding and era.
In abstract, the principle contributions of this paper are as follows:
- Present-O is a unified mannequin that integrates multimodal understanding and era utilizing a single transformer.
- Present-O unifies autoregressive and discrete diffusion modeling inside one transformer, dealing with each textual content and pictures successfully.
- The Present-O framework outperforms or matches particular person baseline fashions with equal or bigger parameters throughout multimodal understanding and era benchmarks.
- Present-O helps downstream functions like text-based inpainting and extrapolation with out fine-tuning and demonstrates potential for mixed-modality era.
- Present-O explores the affect of various kinds of representations, offering beneficial insights for enhancing multimodal understanding in unified fashions.
In recent times, an growing variety of research have centered on unified multimodal language fashions able to each comprehension and era. Some efforts use steady representations interleaved with textual content tokens for autoregressive modeling to generate photographs. SEED-X proposes a unified and versatile basis system able to dealing with each multimodal understanding and era duties. On this strategy, steady picture representations from the CLIP ViT encoder are mixed with textual content tokens and fed into a big language mannequin (LLM) to carry out next-word prediction and picture illustration regression. Chameleon introduces a household of token-based mixed-modal fashions able to each comprehending and producing photographs. This strategy represents all modalities as discrete tokens, using a unified transformer-based structure and coaching the mannequin from scratch in an end-to-end method. Compared, Present-O additionally adopts discrete tokens to symbolize all modalities however makes use of a discrete diffusion course of as a substitute of autoregressive modeling for visible era.
SHOW-O: Methodology and Structure
The first goal behind the Present-O framework is to develop a unified mannequin that integrates autoregressive and diffusion modeling for joint multimodal understanding and era. Growing such a unified mannequin poses vital challenges, with core points revolving round: i) defining the mannequin’s enter/output area; ii) unifying varied forms of enter knowledge from completely different modalities; iii) integrating each autoregressive and diffusion modeling right into a single transformer; and iv) successfully coaching such a unified mannequin.
Present-O addresses these challenges with the next options:
- Present-O constructs the enter/output area by tokenizing textual content and picture knowledge into discrete tokens.
- Present-O introduces its default structure and a unified prompting technique to construction enter knowledge and modalities.
- Present-O demonstrates find out how to incorporate each autoregressive and diffusion modeling inside a single transformer.
- Present-O presents a three-stage coaching pipeline to successfully practice the unified mannequin.
Tokenization
Provided that the proposed Present-O is constructed upon pre-trained LLMs, it’s pure to carry out unified studying within the discrete area. By sustaining a unified vocabulary that features discrete textual content and picture tokens, Present-O is tasked with the identical studying goal: predicting discrete tokens.
Textual content Tokenization
Present-O relies on a pre-trained LLM, and the identical tokenizer is used for textual content knowledge tokenization with none modifications.
Picture Tokenization
Following MAGVIT-v2, Present-O trains a lookup-free quantizer utilizing round 35M picture knowledge. The quantizer maintains a codebook of dimension 8,192 and encodes photographs of 256×256 decision into 16×16 discrete tokens. MAGVIT-v2 is chosen for its ease of fine-tuning, making it appropriate as a video tokenizer with temporal compression functionality, a side Present-O plans to discover sooner or later. Another strategy is to make use of completely different tokenizers for understanding and era, respectively. Impressed by current research, Present-O additionally extracts steady picture representations from the pre-trained MAGVIT-v2 and CLIP-ViT encoder to discover enhancements in multimodal understanding capabilities.. Within the following sections, the default Present-O employs discrete picture tokens as enter for each multimodal understanding and era. For simplicity, the methodology sections will elaborate solely on the default Present-O.
Structure
Present-O inherits the structure of current LLMs with none structure modifications, aside from prepending a QK-Norm operation to every consideration layer. Present-O is initialized with the weights of a pre-trained LLM and expands the dimensions of the embedding layer by incorporating 8,192 new learnable embeddings for discrete picture tokens. In contrast to state-of-the-art diffusion fashions that require an extra textual content encoder, Present-O inherently encodes textual content conditional data for text-to-image era.
Unified Prompting
To carry out unified studying on multimodal understanding and era, Present-O makes use of a unified prompting technique to format varied sorts of enter knowledge. Given an image-text pair (x, y), it’s first tokenized into M picture tokens and N textual content tokens by the picture and textual content tokenizers, respectively. The tokens are then fashioned into an enter sequence in accordance with the duty sort, as illustrated within the following determine.
By using this immediate design, Present-O can successfully encode varied enter knowledge for multimodal understanding, text-to-image era, and mixed-modality era as sequential knowledge. This setup allows unified studying to function seamlessly throughout sequences for these varied duties. As soon as skilled, Present-O may be prompted to deal with a variety of vision-language duties, together with visible query answering and text-to-image era.
Omni-Consideration Mechanism
In contrast to current works that mannequin sequences autoregressively solely, Present-O introduces an omni-attention mechanism, enabling it to mannequin varied forms of alerts in distinct methods. This complete consideration mechanism adaptively switches between causal and full consideration based mostly on the format of the enter sequence. The next determine illustrates examples of omni-attention for various enter sequences.
Particularly, Present-O processes textual content tokens throughout the sequence through causal consideration, whereas picture tokens are dealt with utilizing full consideration, permitting every token to comprehensively work together with all others. In multimodal understanding, textual content tokens can attend to all earlier picture tokens, whereas in text-to-image era, picture tokens can work together with all previous textual content tokens. Omni-attention retains the textual content reasoning information from the pre-trained LLM and enhances the effectivity of picture era by decreasing sampling steps. Moreover, it helps varied downstream functions, similar to inpainting and extrapolation, with out requiring fine-tuning. When given solely textual content tokens, the mechanism defaults to causal consideration.
SHOW-O: Experiments and Outcomes
The next desk presents the multimodal understanding functionality of Present-O on public benchmarks, similar to picture captioning and visible question-answering duties.
The present model of Present-O is constructed upon Phi-1.5, and subsequently, Present-O’s understanding-only counterpart, LLaVA-v1.5-Phi-1.5, serves because the direct baseline. Present-O displays comparable efficiency in all analysis metrics to the baseline LLaVA-v1.5-Phi-1.5, which is devoted solely to multimodal understanding. This demonstrates the nice potential of the Present-O framework to unify multimodal understanding and era inside a single transformer. When in comparison with understanding-only fashions like InstructBLIP, Qwen-VL-Chat, and mPLUG-Owl2, Present-O, regardless of having a a lot smaller mannequin dimension, achieves aggressive efficiency on the POPE, MME, Flickr30k, and VQAv2 benchmarks, and performs higher on the GQA benchmark. When in comparison with unified fashions with considerably extra parameters, similar to NExT-GPT-13B and Chameleon-34B, Present-O additionally achieves robust efficiency on the Flickr30k benchmark and performs a lot better on the VQAv2 benchmark.
Given these promising outcomes, Present-O is envisioned as a possible next-generation basis mannequin for unifying understanding and era. These outcomes additionally exhibit the potential of scaling Present-O to attain state-of-the-art efficiency.
Qualitative Comparisons
We current qualitative comparisons with diffusion-based fashions, similar to SDv1.5, SDXL, and the autoregressive-based mannequin LlamaGen, alongside unified fashions like LWM and SEED-X, as demonstrated within the following determine.
Present-O demonstrates the power to generate sensible photographs with constant content material described in each quick and lengthy textual content prompts. In comparison with SDv1.5 and LlamaGen, Present-O displays higher visible high quality and stronger image-text alignment. For example, within the second column, each SDv1.5 and LlamaGen fail to completely comprehend the textual content immediate and miss attributes similar to “sunset” and “blue domes” within the generated photographs. Compared to SDXL, Present-O offers comparable visible high quality and alignment, as seen in examples like “a rally car race” and “stunning contrast against the vibrant sunset.”
Textual content-Guided Inpainting and Extrapolation
Present-O naturally helps text-based inpainting and extrapolation with out requiring any fine-tuning. The next determine illustrates a number of examples.
On the high of the determine, given an enter picture and an inpainting masks, Present-O can rework a purple trolley automotive right into a blue sports activities automotive with modern curves and tinted home windows based mostly on a user-provided textual content immediate. Present-O may also extrapolate the unique picture horizontally or vertically based mostly on the given textual content immediate. For example, within the second row, Present-O extrapolates a picture by including new objects, like “red wildflowers.” The pixels in each the in-painted and extrapolated areas stay in step with the unique picture. These examples clearly exhibit the inherent benefits of Present-O over autoregressive fashions for downstream functions.
Last Ideas
On this article we have now talked about Present-O, a unified transformer that integrates multimodal understanding and era. In contrast to absolutely autoregressive fashions, Present-O unifies autoregressive and discrete diffusion modeling to adaptively deal with inputs and outputs of assorted and blended modalities. The unified mannequin flexibly helps a variety of vision-language duties, together with visible query answering, text-to-image era, text-guided inpainting/extrapolation, and mixed-modality era. Throughout varied benchmarks, Present-O demonstrates comparable or superior efficiency to current particular person fashions with an equal or bigger variety of parameters, highlighting its potential as a next-generation basis mannequin. On this framework, the mannequin is tasked with predicting Gaussian noise added to the continual latent representations. In distinction, different fashions like D3PM, Masks-predict, ARDM, and MaskGIT use a discrete corruption course of as a substitute for Gaussian diffusion. Present-O is the primary to unify autoregressive and discrete diffusion modeling, enabling it to deal with completely different modalities in distinct methods. Intensive experimental outcomes exhibit that Present-O is similar to, and even higher than, particular person knowledgeable fashions throughout a variety of vision-language duties. This highlights its potential as a next-generation basis mannequin.
