Generalizing diffusion modeling to multimodal, multitask settings

A novel loss function and a way to aggregate multimodal input data are key to dramatic improvements on some test data.

One of the lessons of the machine learning revolution has been that, perhaps counterintuitively, training a model on multiple data types or multiple tasks can improve performance relative to single-purpose models. A model trained on multiple languages, for instance, can learn distinctions that are subtle in one language but pronounced in another, and a model trained on, say, object segmentation may learn properties of visual scenes that help it with depth perception.

Related content
First model to work across a wide range of products uses a second U-Net encoder to capture fine-grained product details.

The advantages of multitask and multimodal training, however, are relatively unexplored in the context of diffusion models, which are responsible for some of the most impressive recent results in generative AI. Diffusion models are trained to incrementally denoise samples to which noise has been incrementally added. The result is that feeding them random noisy inputs will yield randomized outputs that are semantically coherent.

In a paper we presented at the International Conference on Learning Representations (ICLR), we describe a general approach to building multimodal, multitask diffusion models. On the input side, we use modality-specific encoders to map data to a shared diffusion space; on the output side, we use multiple task-specific decoders to map general representations to specific outputs.

MM:MT diffusion architecture.png
The architecture of the multimodal, multitask diffusion model.

The paper presents a theoretical analysis of the problem of generalizing diffusion models to the multimodal, multitask setting, and on the basis of that analysis, it proposes several modifications of the loss function typically used for diffusion modeling. In experiments, we tested our approach on four different multimodal or multitask data sets, and across the board, it was able to match or improve performance relative to single-purpose models.

Minding modality

In the standard diffusion modeling scenario, the model’s encoder maps inputs to a representational space; within that space, a forward process iteratively adds noise to the input representation, and a reverse process iteratively removes it.

Related content
Diffusion modeling within the representational space of a variational autoencoder enables state-of-the-art results.

The loss function includes two terms that measure the distance between the probability distribution of the forward process and the learned probability distribution of the reverse process. One term compares the marginal distributions for the two processes in the forward direction: that is, it compares the likelihoods that any given noisy representation will occur during the forward process. The other term compares the posterior representations of the reverse process — that is, the likelihood that a given representation at time t-1 preceded the representation at time t. We modify these terms so that the distributions are conditioned on the modality of the data — that is, the distributions can differ for data of different modalities.

Both of these loss terms operate in the representational space: they consider the likelihood of a particular representation given another representation. But we also have a term in the loss function that looks at the probability that an input of a given modality led to a particular representation. This helps ensure that the reverse process will correctly recover the modality of the data.

MM:MT diffusion loss.png
The loss function for the multimodal-, multitask diffusion model is the sum of four sublosses, L0–L3. L0 compares the noise distributions of the forward and reverse processes, conditioned on the input data (X). L1 compares posterior distributions, also conditioned on the input data. L2 is the new term in our setting, which induces the model to recover input modalities.

Multimodal means

To fuse the multimodal information used to train the model, we consider the transition distribution in the forward direction, which determines how much noise to add to a given data representation. To compute the mean of that distribution, we define a weighted average of the multimodal input encodings, where the weights are based on input modality.

The transition probability of the forward process. The probability of z sub t, conditioned on z sub t minus 1 and X (the input data) is set equal to a normal distribution whose mean is defined by z sub t minus 1 plus the weighted sum of the encodings of the inputs, sorted by modality. The variance is 1 minus a fraction consisting of a time-varying weight over N (the number of different modalities) plus 1.
The equation for computing the mean and variance of the transition probability of the forward process in the multimodal, multitask setting. N is the number of modalities; wt(i) are the weights assigned to different modalities; xi is the input data; and Ei is the input encoder.

On the basis of the transition probabilities of the forward process, we can now compute the marginal distributions of noisy representations and the posterior distributions of the reverse process (corresponding to sublosses L0 and L1 in the loss function):

The equation for the marginal distribution. The probability of z sub t, conditioned on z sub zero and X (the input data), is set equal to the normal distribution whose mean is the sum of z sub zero (with a coefficient) and a weighted sum of input encoding, sorted by modality. The variance includes a time-varying term (1 minus a time-varying variable), which increases the noise at each time step.
The marginal distribution for the noisy representation zt in the multitask setting (corresponding to subloss L0, above).
The equation for the posterior mean includes a noisy data representation (z sub t), modified by constant factors, from which is subtracted t weighted sum of the encodings of input data of different modalities  (E sub i of x sub i).
The equation for the mean of the posterior distribution, in the multitask setting.

Evaluation

We tested our approach on four tasks, two of which were multitask, and two of which were multimodal. The multitask experiments were both in the vision domain: one involved jointly generating visual data and the associated segmentation masks, and the other was a novel multitask pretraining task in which a diffusion generation model also learned fill in masked regions of input images.

Related content
Generative AI supports the creation, at scale, of complex, realistic driving scenarios that can be directed to specific locations and environments.

The multimodal experiments involved images and other modalities. In one, the model was trained to jointly generate images and their labels, and in the other, the model learned to jointly generate images and their embeddings in a representational space — for instance, CLIP embeddings.

The image segmentation was and embedding generation tasks were chiefly intended as qualitative demonstrations. But the masked pretraining task and the joint generation of images and labels allowed for quantitative evaluation.

Two sets of three images each. In both sets, the first image is of a street scene; the second image is the target segmentation, with objects in the scene masked out in different colors; and the third is the segmentation generated by the model, which is essentially indistinguishable from the target.
Qualitative examples of the segmentation mask generation tasks, with the source image (left), the ground truth segmentation (center), and the masks generated by our method.

We evaluated the masked pretraining model on the task of reconstructing the masked image regions, using learned perceptual image patch similarity (LPIPS) as a metric. LPIPS measures the similarity between two images according to their activations of selected neurons within an image recognition model. Our approach dramatically outperformed the baselines, which were trained only on the reconstruction task, not (simultaneously) on the diffusion task. In some cases, our model’s error rate was almost an order of magnitude lower than the baseline models’.

Two sets of three images each, including a source image, the same image with several black squares of fixed size randomly superimposed upon it, and the model's reconstruction of the complete image.
Our model’s re-creations of masked image regions.

On the task of jointly generating images and labels, our model’s performance was comparable to that of the best baseline vision-language model, with slightly higher precision and slightly lower recall.

For these initial experiments, we evaluated multitask and multimodal performance separately, and each experiment involved only two modalities or tasks. But at least prospectively, the power of our model lies in its generalizability, and in ongoing work, we are evaluating on more than two modalities or tasks at a time and on simultaneous multimodal and multitask training. We are eager to see the result.

Related content

US, MA, N.reading
Amazon Industrial Robotics is seeking exceptional talent to help develop the next generation of advanced robotics systems that will transform automation at Amazon's scale. We're building revolutionary robotic systems that combine cutting-edge AI, sophisticated control systems, and advanced mechanical design to create adaptable automation solutions capable of working safely alongside humans in dynamic environments. This is a unique opportunity to shape the future of robotics and automation at an unprecedented scale, working with world-class teams pushing the boundaries of what's possible in robotic dexterous manipulation, locomotion, and human-robot interaction. We are seeking a talented Applied Scientist to join our advanced robotics team, focusing on developing and applying cutting-edge simulation methodologies for advanced robotics systems. This role centers on research and development of physics-based simulation techniques, sim-to-real transfer methods, and machine learning approaches that enable rapid development, testing, and validation of robotic systems operating in complex, real-world environments. Key job responsibilities - Advance physics-based simulation fidelity for contact-rich manipulation and locomotion - Design and build high-performance simulation tools integrated into a production robotics stack - Translate research ideas into robust, scalable software pipelines - Develop methods to quantify and reduce simulation-to-reality gaps across design, safety, and control - Architect scalable simulation solutions for rigid and deformable body dynamics - Build simulation pipelines optimized for large-scale reinforcement and policy learning - Establish frameworks for continuous simulation improvement using real-world deployment data - Collaborate with engineering, science, and safety teams on simulation requirements and validation About the team Our team is building a comprehensive simulation platform for advanced robotics development, combining locomotion and manipulation capabilities. We operate at the cutting edge of physics simulation, reinforcement learning, and sim-to-real transfer, collaborating with world-class robotics engineers, applied scientists, and mechanical designers in a fast-paced, innovation-driven environment. This role uniquely combines fundamental research with real-world deployment. You will pursue core research questions in physics-based simulation while seeing your work translated into production systems, validated on real hardware, and informed by deployment data. Working alongside Simulation Software Engineers, you will help transform research ideas into scalable, production-grade simulation capabilities that directly impact how robots are designed, trained, and deployed.
US, WA, Redmond
Amazon Leo is Amazon’s low Earth orbit satellite network. Our mission is to deliver fast, reliable internet connectivity to customers beyond the reach of existing networks. From individual households to schools, hospitals, businesses, and government agencies, Amazon Leo will serve people and organizations operating in locations without reliable connectivity. Export Control Requirement: Due to applicable export control laws and regulations, candidates must be a U.S. citizen or national, U.S. permanent resident (i.e., current Green Card holder), or lawfully admitted into the U.S. as a refugee or granted asylum. This position is part of the Satellite Attitude Determination and Control team. You will design and analyze the control system and algorithms, support development of our flight hardware and software, help integrate the satellite in our labs, participate in flight operations, and see a constellation of satellites flow through the production line into orbit. Key job responsibilities - Design and analyze algorithms for estimation, flight control, and precise pointing using linear methods and simulation. - Develop and apply models and simulations, with various levels of fidelity, of the satellite and our constellation. - Component level environmental testing, functional and performance checkout, subsystem integration, satellite integration, and in space operations. - Manage the spacecraft constellation as it grows and evolves. - Continuously improve our ability to serve customers by maximizing payload operations time. - Develop autonomy for Fault Detection and Isolation on board the spacecraft. A day in the life This is an opportunity to play a significant role in the design of an entirely new satellite system with challenging performance requirements. The large, integrated constellation brings opportunities for advanced capabilities that need investigation and development. The constellation size also puts emphasis on engineering excellence so our tools and methods, from conceptualization through manufacturing and all phases of test, will be state of the art as will the satellite and supporting infrastructure on the ground. You will find that Amazon Leo's mission is compelling, so our program is staffed with some of the top engineers in the industry. Our daily collaboration with other teams on the program brings constant opportunity for discovery, learning, and growth. About the team Our team has lots of experience with various satellite systems and many other flight vehicles. We have bench strength in both our mission and core GNC disciplines. We design, prototype, test, iterate and learn together. Because GNC is central to safe flight, we tend to drive Concepts of Operation and many system level analyses.
US, NY, New York
Advertising at Amazon is growing incredibly fast and we are responsible for defining and delivering a collection of advertising products that drive discovery and sales. Amazon Business Ads is equally growing fast ($XXXMs to $XBs) and owns engineering and science for the AB WW ad experience. We build business-to-business (“B2B”) specific ad solutions distributed across retail and ad systems for shopper and advertiser experiences. Some include new ad placements or widgets, creatives, sourcing techniques, ad campaign management capabilities and much more! We consider unique AB qualities which are differentiated from the consumer experience such as varying shopper role types, purchasing complexities based on business size and industry (eg education vs healthcare), AB specific features (eg business discounts, buying policies to restrict and prefer products), and AB buyer behaviors (eg buying in bulk). We are seeking a scientific leader who can drive innovation in complex problem areas and new business initiatives. The ideal candidate will: Technical & Research Requirements: * Demonstrate fluency in Python, R, Matlab or other statistical languages and familiarity with deep learning frameworks like PyTorch, TensorFlow * Lead end-to-end solution development from research to prototyping and experimentation * Write and deploy significant parts of scientifically novel software solutions into production Leadership & Influence: * Drive team's scientific agenda by proposing new initiatives and securing management buy-in including PM, SDM * Build consensus on large projects and influence decisions across different teams in Ads Key Leadership Principles: * Dive Deep: Uncover non-obvious insights in data * Deliver Results: Create solutions aligned with customer and product needs * Learn and Be Curious: Demonstrate self-driven desire to explore new research areas * Earn Trust: Build relationships with stakeholders through understanding business needs
US, WA, Seattle
Prime Video is a first-stop entertainment destination offering customers a vast collection of premium programming in one app available across thousands of devices. Prime members can customize their viewing experience and find their favorite movies, series, documentaries, and live sports – including Amazon MGM Studios-produced series and movies; licensed fan favorites; and programming from Prime Video add-on subscriptions such as Apple TV+, Max, Crunchyroll and MGM+. All customers, regardless of whether they have a Prime membership or not, can rent or buy titles via the Prime Video Store, and can enjoy even more content for free with ads. Are you interested in shaping the future of entertainment? Prime Video's technology teams are creating best-in-class digital video experience. As a Prime Video technologist, you’ll have end-to-end ownership of the product, user experience, design, and technology required to deliver state-of-the-art experiences for our customers. You’ll get to work on projects that are fast-paced, challenging, and varied. You’ll also be able to experiment with new possibilities, take risks, and collaborate with remarkable people. We’ll look for you to bring your diverse perspectives, ideas, and skill-sets to make Prime Video even better for our customers. With global opportunities for talented technologists, you can decide where a career Prime Video Tech takes you! As an Applied Scientist in the Prime Video Playback Intelligence Organization, you will have deep subject matter expertise in applied machine learning and data science, with specializations in video streaming optimization, information retrieval, anomaly detection and root-causing systems, large language models and generative AI across various modalities. Key job responsibilities - Work with multiple teams of scientists, engineers, and product managers to translate business and functional requirements into concrete deliverables leading strategic efforts to enhance customer quality of experiences. - Work on problems spaces such as: improving the customer playback quality of experience across Video on Demand, Live Events and Linear Content. - Reduce the time/cost/effort to optimize the customer experience as well as detect, root-cause, and mitigate defects in the customer experience. You’ll seek to understand the depth and nuance of streaming video at scale and identify opportunities to grow our business and improve customer quality of experience via principled ML/AI solutions. - Lead integration of new algorithms and processes into existing modeling stacks, simplify and streamline the existing modeling stacks, and develop testing and evaluation strategies. Ultimately, you'll work backwards from the desired outcomes and lead the way on determining the ideal solution (statistical techniques, traditional ML, GenAI, etc). A day in the life We love solving challenging and hard problems in our quest to innovate on behalf of our customers and provide the best video streaming experience. We push the boundaries to leverage and invent technologies which help create unrivaled experiences for our customers to help us move fast in a growing and changing environment. We use data to guide our decisions, work closely with our engineering and product counterparts, and partner with other Science teams as well as academic institutions to learn and guide in an environment of innovation.
IN, KA, Bengaluru
Selection Monitoring team is responsible for making the biggest catalog on the planet even bigger. In order to drive expansion of the Amazon catalog, we develop advanced ML/AI technologies to process billions of products and algorithmically find products not already sold on Amazon. We work with structured, semi-structured and Visually Rich Documents using deep learning, NLP and image processing. The role demands a high-performing and flexible candidate who can take responsibility for success of the system and drive solutions from research, prototype, design, coding and deployment. We are looking for Applied Scientists to tackle challenging problems in the areas of Information Extraction, Efficient crawling at internet scale, developing ML models for website comprehension and agents to take multi-step decisions. You should have depth and breadth of knowledge in text mining, information extraction from Visually Rich Documents, semi structured data (HTML) and advanced machine learning. You should also have programming and design skills to manipulate Semi-Structured and unstructured data and systems that work at internet scale. You will encounter many challenges, including: - Scale (build models to handle billions of pages), - Accuracy (requirements for precision and recall) - Speed (generate predictions for millions of new or changed pages with low latency) - Diversity (models need to work across different languages, market places and data sources) You will help us to - Build a scalable system which can algorithmically extract information from world wide web. - Intelligently cluster web pages, segment and classify regions, extract relevant information and structure the data available on semi-structured web. - Build systems that will use existing Knowledge Base to perform open information extraction at scale from visually rich documents. Key job responsibilities - Use AI, NLP and advances in LLMs/SLMs and agentic systems to create scalable solutions for business problems. - Efficiently Crawl web, Automate extraction of relevant information from large amounts of Visually Rich Documents and optimize key processes. - Design, develop, evaluate and deploy, innovative and highly scalable ML models, esp. leveraging latest advances in RL-based fine tuning methods like DPO, GRPO etc. - Work closely with software engineering teams to drive real-time model implementations. - Establish scalable, efficient, automated processes for large scale model development, model validation and model maintenance. - Lead projects and mentor other scientists, engineers in the use of ML techniques. - Publish innovation in research forums.
BR, SP, Sao Paulo
Do you like working on projects that are highly visible and are tied closely to Amazon’s growth? Are you seeking an environment where you can drive innovation leveraging the scalability and innovation with Amazon's AWS cloud services? The Amazon International Technology Team is hiring Applied Scientists to work in our Machine Learning team in Mexico City. The Intech team builds International extensions and new features of the Amazon.com web site for individual countries and creates systems to support Amazon operations. We have already worked in Germany, France, UK, India, China, Italy, Brazil and more. Key job responsibilities About you You want to make changes that help millions of customers. You don’t want to make something 10% better as a part of an enormous team. Rather, you want to innovate with a small community of passionate peers. You have experience in analytics, machine learning, LLMs and Agentic AI, and a desire to learn more about these subjects. You want a trusted role in strategy and product design. You put the customer first in your thinking. You have great problem solving skills. You research the latest data technologies and use them to help you innovate and keep costs low. You have great judgment and communication skills, and a history of delivering results. Your Responsibilities - Define and own complex machine learning solutions in the consumer space, including targeting, measurement, creative optimization, and multivariate testing. - Design, implement, and evolve Agentic AI systems that can autonomously perceive their environment, reason about context, and take actions across business workflows—while ensuring human-in-the-loop oversight for high-stakes decisions. - Influence the broader team's approach to integrating machine learning into business workflows. - Advise leadership, both tech and non-tech. - Support technical trade-offs between short-term needs and long-term goals.
US, WA, Bellevue
Alexa International Science team is looking for a passionate, talented, and inventive Senior Applied Scientist to help build industry-leading technology with Large Language Models (LLMs) and multimodal systems, requiring strong deep learning and generative models knowledge. At this level, you will drive cross-team scientific strategy, influence partner teams, and deliver solutions that have broad impact across Alexa's international products and services. Key job responsibilities As a Senior Applied Scientist with the Alexa International team, you will work with talented peers to develop novel algorithms and modeling techniques to advance the state of the art with LLMs, particularly delivering industry-leading scientific research and applied AI for multi-lingual applications — a challenging area for the industry globally. Your work will directly impact our global customers in the form of products and services that support Alexa+. You will leverage Amazon's heterogeneous data sources and large-scale computing resources to accelerate advances in text, speech, and vision domains. The ideal candidate possesses a solid understanding of machine learning, speech and/or natural language processing, modern LLM architectures, LLM evaluation & tooling, and a passion for pushing boundaries in this vast and quickly evolving field. They thrive in fast-paced environment, like to tackle complex challenges, excel at swiftly delivering impactful solutions while iterating based on user feedback, and are able to influence and align multiple teams around a shared scientific vision.
US, NY, New York
We are looking for detail-oriented, organized, and responsible individuals who are eager to learn how to apply their structural econometrics skillsets to solve real world problems. The intern will work in the area of Amazon Private Brands and develop models to improve our product selection. Our PhD Economist Internship Program offers hands-on experience in applied economics, supported by mentorship, structured feedback, and professional development. Interns work on real business and research problems, building skills that prepare them for full-time economist roles at Amazon and beyond. You will learn how to build data sets and perform applied econometric analysis collaborating with economists, scientists, and product managers. These skills will translate well into writing applied chapters in your dissertation and provide you with work experience that may help you with placement. These are full-time positions at 40 hours per week, with compensation being awarded on an hourly basis. About the team The Amazon Private Brands science advance team applies Machine Learning, Statistics and Econometrics/economics to solve high-impact business problems, develop prototypes for Amazon-scale science solutions, and optimize key business functions of Amazon Private Brands and other Amazon orgs. We are an interdisciplinary team, using science and technology and leveraging the strengths of engineers and scientists to build solutions for some of the toughest business problems at Amazon, covering areas such as pricing, discovery, negotiation, forecasting, supply chain and product selection/development.
US, WA, Seattle
Innovators wanted! Are you an entrepreneur? A builder? A dreamer? This role is part of an Amazon Special Projects team that takes the company’s Think Big leadership principle to the extreme. We focus on creating entirely new products and services with a goal of positively impacting the lives of our customers. No industries or subject areas are out of bounds. If you’re interested in innovating at scale to address big challenges in the world, this is the team for you. Here at Amazon, we embrace our differences. We are committed to furthering our culture of inclusion. We have thirteen employee-led affinity groups, reaching 40,000 employees in over 190 chapters globally. We are constantly learning through programs that are local, regional, and global. Amazon’s culture of inclusion is reinforced within our 16 Leadership Principles, which remind team members to seek diverse perspectives, learn and be curious, and earn trust. Our team highly values work-life balance, mentorship and career growth. We believe striking the right balance between your personal and professional life is critical to life-long happiness and fulfillment. We care about your career growth and strive to assign projects and offer training that will challenge you to become your best.
US, CA, San Francisco
Amazon has launched a new research lab in San Francisco to develop foundational capabilities for useful AI agents. We’re enabling practical AI to make our customers more productive, empowered, and fulfilled. Our work leverages large vision language models (VLMs) with reinforcement learning (RL) and world modeling to solve perception, reasoning, and planning to build useful enterprise agents. Our lab is a small, talent-dense team with the resources and scale of Amazon. Each team in the lab has the autonomy to move fast and the long-term commitment to pursue high-risk, high-payoff research. We’re entering an exciting new era where agents can redefine what AI makes possible. Key job responsibilities You will contribute directly to AI agent development in an applied research role to improve the multi-model perception and visual-reasoning abilities of our agent. Daily responsibilities including model training, dataset design, and pre- and post-training optimization. You will be hired as a Member of Technical Staff.