Daily Papers

This paper presents Dolphin, a novel decoder-decoder architecture for energy-efficient processing of long contexts in language models. Our approach addresses the significant energy consumption and latency challenges inherent in on-device models. Dolphin employs a compact 0.5B parameter decoder to distill extensive contextual information into a memory embedding, substantially reducing the input length for the primary 7B parameter decoder model. Inspired by vision-language models, we repurpose the image embedding projector to encode long textual contexts, effectively treating extended context as a distinct modality. This innovative method enables processing of substantially longer contexts without the typical computational overhead associated with extended input sequences. Empirical evaluations demonstrate a 10-fold improvement in energy efficiency and a 5-fold reduction in latency compared to conventional full-length context processing methods without losing quality of the response. Our work contributes to the development of more sustainable and scalable language models for on-device applications, addressing the critical need for energy-efficient and responsive AI technologies in resource-constrained environments while maintaining the accuracy to understand long contexts. This research has implications for the broader field of natural language processing, particularly in the domain of efficient model design for resource-limited settings. By enabling more sophisticated AI capabilities on edge devices, Dolphin paves the way for advanced language processing in a wide range of applications where computational resources are at a premium. The Dolphin model is publicly available at https://huggingface.co/NexaAIDev/Dolphin.

While large-scale pre-trained text-to-image models can synthesize diverse and high-quality human-centric images, an intractable problem is how to preserve the face identity for conditioned face images. Existing methods either require time-consuming optimization for each face-identity or learning an efficient encoder at the cost of harming the editability of models. In this work, we present an optimization-free method for each face identity, meanwhile keeping the editability for text-to-image models. Specifically, we propose a novel face-identity encoder to learn an accurate representation of human faces, which applies multi-scale face features followed by a multi-embedding projector to directly generate the pseudo words in the text embedding space. Besides, we propose self-augmented editability learning to enhance the editability of models, which is achieved by constructing paired generated face and edited face images using celebrity names, aiming at transferring mature ability of off-the-shelf text-to-image models in celebrity faces to unseen faces. Extensive experiments show that our methods can generate identity-preserved images under different scenes at a much faster speed.

Large-scale foundation models, such as CLIP, have demonstrated remarkable success in visual recognition tasks by embedding images in a semantically rich space. Self-supervised learning (SSL) has also shown promise in improving visual recognition by learning invariant features. However, the combination of CLIP with SSL is found to face challenges due to the multi-task framework that blends CLIP's contrastive loss and SSL's loss, including difficulties with loss weighting and inconsistency among different views of images in CLIP's output space. To overcome these challenges, we propose a prompt learning-based model called GOPro, which is a unified framework that ensures similarity between various augmented views of input images in a shared image-text embedding space, using a pair of learnable image and text projectors atop CLIP, to promote invariance and generalizability. To automatically learn such prompts, we leverage the visual content and style primitives extracted from pre-trained CLIP and adapt them to the target task. In addition to CLIP's cross-domain contrastive loss, we introduce a visual contrastive loss and a novel prompt consistency loss, considering the different views of the images. GOPro is trained end-to-end on all three loss objectives, combining the strengths of CLIP and SSL in a principled manner. Empirical evaluations demonstrate that GOPro outperforms the state-of-the-art prompting techniques on three challenging domain generalization tasks across multiple benchmarks by a significant margin. Our code is available at https://github.com/mainaksingha01/GOPro.

This paper presents a powerful framework to customize video creations by incorporating multiple specific identity (ID) photos, with video diffusion Transformers, referred to as Ingredients. Generally, our method consists of three primary modules: (i) a facial extractor that captures versatile and precise facial features for each human ID from both global and local perspectives; (ii) a multi-scale projector that maps face embeddings into the contextual space of image query in video diffusion transformers; (iii) an ID router that dynamically combines and allocates multiple ID embedding to the corresponding space-time regions. Leveraging a meticulously curated text-video dataset and a multi-stage training protocol, Ingredients demonstrates superior performance in turning custom photos into dynamic and personalized video content. Qualitative evaluations highlight the advantages of proposed method, positioning it as a significant advancement toward more effective generative video control tools in Transformer-based architecture, compared to existing methods. The data, code, and model weights are publicly available at: https://github.com/feizc/Ingredients.

This paper proposes a pre-trained neural network for handling event camera data. Our model is a self-supervised learning framework, and uses paired event camera data and natural RGB images for training. Our method contains three modules connected in a sequence: i) a family of event data augmentations, generating meaningful event images for self-supervised training; ii) a conditional masking strategy to sample informative event patches from event images, encouraging our model to capture the spatial layout of a scene and accelerating training; iii) a contrastive learning approach, enforcing the similarity of embeddings between matching event images, and between paired event and RGB images. An embedding projection loss is proposed to avoid the model collapse when enforcing the event image embedding similarities. A probability distribution alignment loss is proposed to encourage the event image to be consistent with its paired RGB image in the feature space. Transfer learning performance on downstream tasks shows the superiority of our method over state-of-the-art methods. For example, we achieve top-1 accuracy at 64.83% on the N-ImageNet dataset.

Composed Image Retrieval (CIR) aims to retrieve a target image based on a reference image and conditioning text, enabling controllable searches. Due to the expensive dataset construction cost for CIR triplets, a zero-shot (ZS) CIR setting has been actively studied to eliminate the need for human-collected triplet datasets. The mainstream of ZS-CIR employs an efficient projection module that projects a CLIP image embedding to the CLIP text token embedding space, while fixing the CLIP encoders. Using the projected image embedding, these methods generate image-text composed features by using the pre-trained text encoder. However, their CLIP image and text encoders suffer from the task discrepancy between the pre-training task (text leftrightarrow image) and the target CIR task (image + text leftrightarrow image). Conceptually, we need expensive triplet samples to reduce the discrepancy, but we use cheap text triplets instead and update the text encoder. To that end, we introduce the Reducing Task Discrepancy of text encoders for Composed Image Retrieval (RTD), a plug-and-play training scheme for the text encoder that enhances its capability using a novel target-anchored text contrastive learning. We also propose two additional techniques to improve the proposed learning scheme: a hard negatives-based refined batch sampling strategy and a sophisticated concatenation scheme. Integrating RTD into the state-of-the-art projection-based ZS-CIR methods significantly improves performance across various datasets and backbones, demonstrating its efficiency and generalizability.

Style-guided texture generation aims to generate a texture that is harmonious with both the style of the reference image and the geometry of the input mesh, given a reference style image and a 3D mesh with its text description. Although diffusion-based 3D texture generation methods, such as distillation sampling, have numerous promising applications in stylized games and films, it requires addressing two challenges: 1) decouple style and content completely from the reference image for 3D models, and 2) align the generated texture with the color tone, style of the reference image, and the given text prompt. To this end, we introduce StyleTex, an innovative diffusion-model-based framework for creating stylized textures for 3D models. Our key insight is to decouple style information from the reference image while disregarding content in diffusion-based distillation sampling. Specifically, given a reference image, we first decompose its style feature from the image CLIP embedding by subtracting the embedding's orthogonal projection in the direction of the content feature, which is represented by a text CLIP embedding. Our novel approach to disentangling the reference image's style and content information allows us to generate distinct style and content features. We then inject the style feature into the cross-attention mechanism to incorporate it into the generation process, while utilizing the content feature as a negative prompt to further dissociate content information. Finally, we incorporate these strategies into StyleTex to obtain stylized textures. The resulting textures generated by StyleTex retain the style of the reference image, while also aligning with the text prompts and intrinsic details of the given 3D mesh. Quantitative and qualitative experiments show that our method outperforms existing baseline methods by a significant margin.

Encouraged by the growing availability of pre-trained 2D diffusion models, image-to-3D generation by leveraging Score Distillation Sampling (SDS) is making remarkable progress. Most existing methods combine novel-view lifting from 2D diffusion models which usually take the reference image as a condition while applying hard L2 image supervision at the reference view. Yet heavily adhering to the image is prone to corrupting the inductive knowledge of the 2D diffusion model leading to flat or distorted 3D generation frequently. In this work, we reexamine image-to-3D in a novel perspective and present Isotropic3D, an image-to-3D generation pipeline that takes only an image CLIP embedding as input. Isotropic3D allows the optimization to be isotropic w.r.t. the azimuth angle by solely resting on the SDS loss. The core of our framework lies in a two-stage diffusion model fine-tuning. Firstly, we fine-tune a text-to-3D diffusion model by substituting its text encoder with an image encoder, by which the model preliminarily acquires image-to-image capabilities. Secondly, we perform fine-tuning using our Explicit Multi-view Attention (EMA) which combines noisy multi-view images with the noise-free reference image as an explicit condition. CLIP embedding is sent to the diffusion model throughout the whole process while reference images are discarded once after fine-tuning. As a result, with a single image CLIP embedding, Isotropic3D is capable of generating multi-view mutually consistent images and also a 3D model with more symmetrical and neat content, well-proportioned geometry, rich colored texture, and less distortion compared with existing image-to-3D methods while still preserving the similarity to the reference image to a large extent. The project page is available at https://isotropic3d.github.io/. The code and models are available at https://github.com/pkunliu/Isotropic3D.

Text-to-image (T2I) models can effectively capture the content or style of reference images to perform high-quality customization. A representative technique for this is fine-tuning using low-rank adaptations (LoRA), which enables efficient model customization with reference images. However, fine-tuning with a limited number of reference images often leads to overfitting, resulting in issues such as prompt misalignment or content leakage. These issues prevent the model from accurately following the input prompt or generating undesired objects during inference. To address this problem, we examine the text embeddings that guide the diffusion model during inference. This study decomposes the text embedding matrix and conducts a component analysis to understand the embedding space geometry and identify the cause of overfitting. Based on this, we propose DECOR, which projects text embeddings onto a vector space orthogonal to undesired token vectors, thereby reducing the influence of unwanted semantics in the text embeddings. Experimental results demonstrate that DECOR outperforms state-of-the-art customization models and achieves Pareto frontier performance across text and visual alignment evaluation metrics. Furthermore, it generates images more faithful to the input prompts, showcasing its effectiveness in addressing overfitting and enhancing text-to-image customization.

Despite the growing use of transformer models in computer vision, a mechanistic understanding of these networks is still needed. This work introduces a method to reverse-engineer Vision Transformers trained to solve image classification tasks. Inspired by previous research in NLP, we demonstrate how the inner representations at any level of the hierarchy can be projected onto the learned class embedding space to uncover how these networks build categorical representations for their predictions. We use our framework to show how image tokens develop class-specific representations that depend on attention mechanisms and contextual information, and give insights on how self-attention and MLP layers differentially contribute to this categorical composition. We additionally demonstrate that this method (1) can be used to determine the parts of an image that would be important for detecting the class of interest, and (2) exhibits significant advantages over traditional linear probing approaches. Taken together, our results position our proposed framework as a powerful tool for mechanistic interpretability and explainability research.

We propose a new transformer-based image and video tokenizer with Binary Spherical Quantization (BSQ). BSQ projects the high-dimensional visual embedding to a lower-dimensional hypersphere and then applies binary quantization. BSQ is (1) parameter-efficient without an explicit codebook, (2) scalable to arbitrary token dimensions, and (3) compact: compressing visual data by up to 100times with minimal distortion. Our tokenizer uses a transformer encoder and decoder with simple block-wise causal masking to support variable-length videos as input. The resulting BSQ-ViT achieves state-of-the-art visual reconstruction quality on image and video reconstruction benchmarks with 2.4times throughput compared to the best prior methods. Furthermore, by learning an autoregressive prior for adaptive arithmetic coding, BSQ-ViT achieves comparable results on video compression with state-of-the-art video compression standards. BSQ-ViT also enables masked language models to achieve competitive image synthesis quality to GAN- and diffusion-based methods.

Composed image retrieval (CIR) task takes a composed query of image and text, aiming to search relative images for both conditions. Conventional CIR approaches need a training dataset composed of triplets of query image, query text, and target image, which is very expensive to collect. Several recent works have worked on the zero-shot (ZS) CIR paradigm to tackle the issue without using pre-collected triplets. However, the existing ZS-CIR methods show limited backbone scalability and generalizability due to the lack of diversity of the input texts during training. We propose a novel CIR framework, only using language for its training. Our LinCIR (Language-only training for CIR) can be trained only with text datasets by a novel self-supervision named self-masking projection (SMP). We project the text latent embedding to the token embedding space and construct a new text by replacing the keyword tokens of the original text. Then, we let the new and original texts have the same latent embedding vector. With this simple strategy, LinCIR is surprisingly efficient and highly effective; LinCIR with CLIP ViT-G backbone is trained in 48 minutes and shows the best ZS-CIR performances on four different CIR benchmarks, CIRCO, GeneCIS, FashionIQ, and CIRR, even outperforming supervised method on FashionIQ. Code is available at https://github.com/navervision/lincir

Generative domain adaptation has achieved remarkable progress, enabling us to adapt a pre-trained generator to a new target domain. However, existing methods simply adapt the generator to a single target domain and are limited to a single modality, either text-driven or image-driven. Moreover, they are prone to overfitting domain-specific attributes, which inevitably compromises cross-domain consistency. In this paper, we propose UniHDA, a unified and versatile framework for generative hybrid domain adaptation with multi-modal references from multiple domains. We use CLIP encoder to project multi-modal references into a unified embedding space and then linear interpolate the direction vectors from multiple target domains to achieve hybrid domain adaptation. To ensure the cross-domain consistency, we propose a novel cross-domain spatial structure (CSS) loss that maintains detailed spatial structure information between source and target generator. Experiments show that the adapted generator can synthesise realistic images with various attribute compositions. Additionally, our framework is versatile to multiple generators, \eg, StyleGAN2 and Diffusion Models.

Monocular depth estimation is an ill-posed problem as the same 2D image can be projected from infinite 3D scenes. Although the leading algorithms in this field have reported significant improvement, they are essentially geared to the particular compound of pictorial observations and camera parameters (i.e., intrinsics and extrinsics), strongly limiting their generalizability in real-world scenarios. To cope with this challenge, this paper proposes a novel ground embedding module to decouple camera parameters from pictorial cues, thus promoting the generalization capability. Given camera parameters, the proposed module generates the ground depth, which is stacked with the input image and referenced in the final depth prediction. A ground attention is designed in the module to optimally combine ground depth with residual depth. Our ground embedding is highly flexible and lightweight, leading to a plug-in module that is amenable to be integrated into various depth estimation networks. Experiments reveal that our approach achieves the state-of-the-art results on popular benchmarks, and more importantly, renders significant generalization improvement on a wide range of cross-domain tests.

Despite unprecedented ability in imaginary creation, large text-to-image models are further expected to express customized concepts. Existing works generally learn such concepts in an optimization-based manner, yet bringing excessive computation or memory burden. In this paper, we instead propose a learning-based encoder for fast and accurate concept customization, which consists of global and local mapping networks. In specific, the global mapping network separately projects the hierarchical features of a given image into multiple ``new'' words in the textual word embedding space, i.e., one primary word for well-editable concept and other auxiliary words to exclude irrelevant disturbances (e.g., background). In the meantime, a local mapping network injects the encoded patch features into cross attention layers to provide omitted details, without sacrificing the editability of primary concepts. We compare our method with prior optimization-based approaches on a variety of user-defined concepts, and demonstrate that our method enables more high-fidelity inversion and robust editability with a significantly faster encoding process. Our code will be publicly available at https://github.com/csyxwei/ELITE.

In recent years, the fashion industry has increasingly adopted AI technologies to enhance customer experience, driven by the proliferation of e-commerce platforms and virtual applications. Among the various tasks, virtual try-on and multimodal fashion image editing -- which utilizes diverse input modalities such as text, garment sketches, and body poses -- have become a key area of research. Diffusion models have emerged as a leading approach for such generative tasks, offering superior image quality and diversity. However, most existing virtual try-on methods rely on having a specific garment input, which is often impractical in real-world scenarios where users may only provide textual specifications. To address this limitation, in this work we introduce Fashion Retrieval-Augmented Generation (Fashion-RAG), a novel method that enables the customization of fashion items based on user preferences provided in textual form. Our approach retrieves multiple garments that match the input specifications and generates a personalized image by incorporating attributes from the retrieved items. To achieve this, we employ textual inversion techniques, where retrieved garment images are projected into the textual embedding space of the Stable Diffusion text encoder, allowing seamless integration of retrieved elements into the generative process. Experimental results on the Dress Code dataset demonstrate that Fashion-RAG outperforms existing methods both qualitatively and quantitatively, effectively capturing fine-grained visual details from retrieved garments. To the best of our knowledge, this is the first work to introduce a retrieval-augmented generation approach specifically tailored for multimodal fashion image editing.

Though the success of CLIP-based training recipes in vision-language models, their scalability to more modalities (e.g., 3D, audio, etc.) is limited to large-scale data, which is expensive or even inapplicable for rare modalities. In this paper, we present ViT-Lens that facilitates efficient omni-modal representation learning by perceiving novel modalities with a pretrained ViT and aligning to a pre-defined space. Specifically, the modality-specific lens is tuned to project multimodal signals to the shared embedding space, which are then processed by a strong ViT that carries pre-trained image knowledge. The encoded multimodal representations are optimized toward aligning with the modal-independent space, pre-defined by off-the-shelf foundation models. A well-trained lens with a ViT backbone has the potential to serve as one of these foundation models, supervising the learning of subsequent modalities. ViT-Lens provides a unified solution for representation learning of increasing modalities with two appealing benefits: (i) Exploiting the pretrained ViT across tasks and domains effectively with efficient data regime; (ii) Emergent downstream capabilities of novel modalities are demonstrated due to the modality alignment space. We evaluate ViT-Lens in the context of 3D as an initial verification. In zero-shot 3D classification, ViT-Lens achieves substantial improvements over previous state-of-the-art, showing 52.0% accuracy on Objaverse-LVIS, 87.4% on ModelNet40, and 60.6% on ScanObjectNN. Furthermore, we enable zero-shot 3D question-answering by simply integrating the trained 3D lens into the InstructBLIP model without any adaptation. We will release the results of ViT-Lens on more modalities in the near future.

Knowledge is acquired by humans through experience, and no boundary is set between the kinds of knowledge or skill levels we can achieve on different tasks at the same time. When it comes to Neural Networks, that is not the case. The breakthroughs in the field are extremely task and domain-specific. Vision and language are dealt with in separate manners, using separate methods and different datasets. Current text classification methods, mostly rely on obtaining contextual embeddings for input text samples, then training a classifier on the embedded dataset. Transfer learning in Language-related tasks in general, is heavily used in obtaining the contextual text embeddings for the input samples. In this work, we propose to use the knowledge acquired by benchmark Vision Models which are trained on ImageNet to help a much smaller architecture learn to classify text. A data transformation technique is used to create a new image dataset, where each image represents a sentence embedding from the last six layers of BERT, projected on a 2D plane using a t-SNE based method. We trained five models containing early layers sliced from vision models which are pretrained on ImageNet, on the created image dataset for the IMDB dataset embedded with the last six layers of BERT. Despite the challenges posed by the very different datasets, experimental results achieved by this approach which links large pretrained models on both language and vision, are very promising, without employing compute resources. Specifically, Sentiment Analysis is achieved by five different models on the same image dataset obtained after BERT embeddings are transformed into gray scale images. Index Terms: BERT, Convolutional Neural Networks, Domain Adaptation, image classification, Natural Language Processing, t-SNE, text classification, Transfer Learning

Training a Large Visual Language Model (LVLM) from scratch, like GPT-4, is resource-intensive. Our paper presents a play-and-plug module for Large Language Models (LLMs), namely Interactive Perception Network (IPN), aiming to achieve a LVLM by incorporating the image understanding capability into LLMs. Previous methods incorporate visual information into LLMs with a simple visual mapping network, where the image feature is projected into the embedding space of LLMs via a linear layer. Such mapping network projects the image feature once yet does not consider the interaction between the image and the human input query. Hence, the obtained visual information with no connections with human intention may be inadequate for LLMs to make intention-following responses, which we term as static visual information. IPN addresses this issue by allowing the LLM to request the desired visual information aligned with various human instructions, which we term as the dynamic interaction between the LLM and visual information. Specifically, IPN consists of a simple visual mapping network to provide the basic perception of an image for LLMs. It also contains additional modules responsible for acquiring requests from LLMs, performing request-based visual information interaction, and transmitting the resulting interacted visual information to LLMs, respectively. In this way, LLMs act to understand the human query, deliver the corresponding request to the request-based visual information interaction module, and generate the response based on the interleaved multimodal information. We evaluate IPN through extensive experiments on multimodal question answering, reasoning, and so on, demonstrating that it significantly improves the zero-shot performance of LVLMs on various multimodal tasks compared to previous methods.

Microscopic assessment of histopathology images is vital for accurate cancer diagnosis and treatment. Whole Slide Image (WSI) classification and captioning have become crucial tasks in computer-aided pathology. However, microscopic WSI face challenges such as redundant patches and unknown patch positions due to subjective pathologist captures. Moreover, generating automatic pathology captions remains a significant challenge. To address these issues, we introduce a novel GNN-ViTCap framework for classification and caption generation from histopathological microscopic images. First, a visual feature extractor generates patch embeddings. Redundant patches are then removed by dynamically clustering these embeddings using deep embedded clustering and selecting representative patches via a scalar dot attention mechanism. We build a graph by connecting each node to its nearest neighbors in the similarity matrix and apply a graph neural network to capture both local and global context. The aggregated image embeddings are projected into the language model's input space through a linear layer and combined with caption tokens to fine-tune a large language model. We validate our method on the BreakHis and PatchGastric datasets. GNN-ViTCap achieves an F1 score of 0.934 and an AUC of 0.963 for classification, along with a BLEU-4 score of 0.811 and a METEOR score of 0.569 for captioning. Experimental results demonstrate that GNN-ViTCap outperforms state of the art approaches, offering a reliable and efficient solution for microscopy based patient diagnosis.

In recent times there has been a surge of multi-modal architectures based on Large Language Models, which leverage the zero shot generation capabilities of LLMs and project image embeddings into the text space and then use the auto-regressive capacity to solve tasks such as VQA, captioning, and image retrieval. We name these architectures as "bridge-architectures" as they project from the image space to the text space. These models deviate from the traditional recipe of training transformer based multi-modal models, which involve using large-scale pre-training and complex multi-modal interactions through co or cross attention. However, the capabilities of bridge architectures have not been tested on complex visual reasoning tasks which require fine grained analysis about the image. In this project, we investigate the performance of these bridge-architectures on the NLVR2 dataset, and compare it to state-of-the-art transformer based architectures. We first extend the traditional bridge architectures for the NLVR2 dataset, by adding object level features to faciliate fine-grained object reasoning. Our analysis shows that adding object level features to bridge architectures does not help, and that pre-training on multi-modal data is key for good performance on complex reasoning tasks such as NLVR2. We also demonstrate some initial results on a recently bridge-architecture, LLaVA, in the zero shot setting and analyze its performance.

Fine-grained and instance-level recognition methods are commonly trained and evaluated on specific domains, in a model per domain scenario. Such an approach, however, is impractical in real large-scale applications. In this work, we address the problem of universal image embedding, where a single universal model is trained and used in multiple domains. First, we leverage existing domain-specific datasets to carefully construct a new large-scale public benchmark for the evaluation of universal image embeddings, with 241k query images, 1.4M index images and 2.8M training images across 8 different domains and 349k classes. We define suitable metrics, training and evaluation protocols to foster future research in this area. Second, we provide a comprehensive experimental evaluation on the new dataset, demonstrating that existing approaches and simplistic extensions lead to worse performance than an assembly of models trained for each domain separately. Finally, we conducted a public research competition on this topic, leveraging industrial datasets, which attracted the participation of more than 1k teams worldwide. This exercise generated many interesting research ideas and findings which we present in detail. Project webpage: https://cmp.felk.cvut.cz/univ_emb/

CLIP is a discriminative model trained to align images and text in a shared embedding space. Due to its multimodal structure, it serves as the backbone of many generative pipelines, where a decoder is trained to map from the shared space back to images. In this work, we show that image synthesis is nevertheless possible using CLIP alone -- without any decoder, training, or fine-tuning. Our approach optimizes a frequency-aware implicit neural representation that encourages coarse-to-fine generation by stratifying frequencies across network layers. To stabilize this inverse mapping, we introduce adversarially robust initialization, a lightweight Orthogonal Procrustes projection to align local text and image embeddings, and a blending loss that anchors outputs to natural image statistics. Without altering CLIP's weights, this framework unlocks capabilities such as text-to-image generation, style transfer, and image reconstruction. These findings suggest that discriminative models may hold untapped generative potential, hidden in plain sight.

In this paper, we are interested in modeling a how-to instructional procedure, such as a cooking recipe, with a meaningful and rich high-level representation. Specifically, we propose to represent cooking recipes and food images as cooking programs. Programs provide a structured representation of the task, capturing cooking semantics and sequential relationships of actions in the form of a graph. This allows them to be easily manipulated by users and executed by agents. To this end, we build a model that is trained to learn a joint embedding between recipes and food images via self-supervision and jointly generate a program from this embedding as a sequence. To validate our idea, we crowdsource programs for cooking recipes and show that: (a) projecting the image-recipe embeddings into programs leads to better cross-modal retrieval results; (b) generating programs from images leads to better recognition results compared to predicting raw cooking instructions; and (c) we can generate food images by manipulating programs via optimizing the latent code of a GAN. Code, data, and models are available online.