Post-Doctorant F/H Editing and Conditional Generation with Text-to-Video Generation Models
Post-Doctorant F/H Editing and Conditional Generation with Text-to-Video Generation Models
Fonction : Post-Doctorant
Titre : Editing and Conditional Generation with Text-to-Video Generation Models
Funding : BPI contract
Contexte :
Recent advancements in generative AI, and in particular diffusion models (1,2), have significantly enhanced the capabilities of text-to-video (T2V) models (3,4), allowing users to produce richly varied and imaginative scenes from natural language descriptions. These systems demonstrate strong scene diversity and flexibility, making them attractive for applications in entertainment, simulation, and human-computer interaction.
However, a persistent limitation lies in their inability to enforce fine-grained conditioning and maintain strict consistency for specific visual elements. For example, while a T2V model can generate a "person walking in a park", it struggles to ensure the persistent appearance of a specific object, character identity, or detailed attribute (such as a specific garment (5)) across complex poses and dynamic environmental interactions.
In contrast, highly specialized image and video editing systems - such as those designed for virtual try-on (5), face swapping, or precise object insertion - excel at fine-grained conditioning on target individuals or objects. They can adapt elements to morphology, pose, and texture details with remarkable realism. Yet, these specialized approaches generally operate in isolation, lacking the scene diversity and broader contextual awareness that foundational T2V models offer.
Bridging these two paradigms offers a powerful opportunity: to synthesize realistic, precisely controllable subjects and objects embedded within richly described, dynamic environments. To achieve this, novel alignment and editing techniques are required. Specifically, post-training with Reinforcement Learning (RL) presents a highly promising methodology to overcome these limitations. By leveraging RL during the post-training phase, foundation T2V models can be explicitly optimized to follow complex conditioning signals, enforce temporal consistency, and align with specific human-defined objectives for fine-grained editing tasks without sacrificing their generative diversity.
.
Research Objectives : The primary mission of the Postdoctoral Research Fellow will be to advance the state-of-the-art in controllable and editable Text-to-Video (T2V) generation. The successful candidate will design, implement, and evaluate novel deep generative models and methodologies that address the current limitations of existing T2V systems. A core focus will be on achieving fine-grained conditional generation via post-training with Reinforcement Learning (RL), allowing users to specify complex temporal, spatial, and stylistic constraints, as well as enabling intuitive and high-fidelity post-generation editing of the video content. The research will aim to produce models that are not only photorealistic but also exhibit high semantic fidelity, temporal coherence, and practical usability in creative and industrial applications.
2. Main Tasks
The Postdoctoral Research Fellow will be responsible for the following main tasks. They will engage in Model Design and Development by designing and implementing novel architectures (e.g., Diffusion Models, Transformers, VAEs) specifically tailored for high-resolution, temporally consistent, and controllable video generation. A key focus is to develop conditional generation techniques to guide the Text-to-Video process using various complex inputs beyond a simple text prompt, such as image references, motion skeletons, semantic masks, or detailed scene descriptions. They will extensively research Video Editing and Manipulation , developing methods for high-fidelity post-generation video editing , allowing for non-destructive modification of generated videos (e.g., object replacement, style transfer, background alteration) while maintaining strong temporal consistency. Furthermore, they will investigate in-context editing mechanisms that enable precise changes to specific segments or objects within a generated video based on new text or image prompts. A core part of the role is Addressing Key T2V Challenges . This includes tackling the fundamental challenge of temporal coherence and consistency , ensuring that generated videos do not suffer from "flickering" or object identity changes across frames, and developing strategies to improve semantic fidelity , resolving issues where models misinterpret complex text prompts. They will also explore methods for efficient training and inference to manage the significant computational cost associated with high-resolution, long-duration video generation, and address the difficulties of data scarcity and bias through techniques like data augmentation or cross-modal transfer learning. Finally, they will perform Evaluation and Benchmarking , establishing rigorous quantitative and qualitative metrics to assess the quality, editability, and controllability of the developed models. The fellow is expected to prioritize Dissemination and Collaboration , which involves documenting research findings and publishing high-quality papers in top-tier machine learning and computer vision venues, actively participating in departmental seminars and contributing to collaborative projects.
Compétences techniques et niveau requis :We are seeking a motivated PhD candidate with a strong background in one or more the following areas :
- speech processing, computer vision, machine learning,
- interest in connecting AI with human cognition Prior experience with LLM, SpeechLMs, RL algorithms, or robotic platforms is a plus, but not mandatory