New opportunity

AI Researcher - Content Video Generation(Post training)

Shopee · Singapore

About this role

Shopee is hiring an AI Researcher in Singapore for Sea Group’s AI department. The role focuses on post-training generative video models for high-fidelity visual synthesis, with applications in social and e-commerce platforms.

Responsibilities

Lead the architecture and implementation of video generation post-training, emphasizing high-quality instruction data, preference alignment and improved video quality.
Develop capabilities for long-video modeling, storyline consistency and precise camera control.
Build video quality evaluation pipelines and multidimensional reward models. Conduct alignment training using RLHF, DPO, GRPO and PPO.
Research and implement model distillation and other methods to accelerate diffusion-model inference.

Requirements

Master’s or PhD in Computer Science or a related field; a bachelor’s degree may be considered with strong industrial experience.
Strong theoretical and practical grounding in video generation post-training, including preference alignment, fine-tuning with LoRA, QLoRA or DoRA, and distillation approaches such as Consistency Models and Flow Matching.
Demonstrated video generation post-training work, including practical experience with video reward models and quality evaluation metrics. Excellent communication and teamwork skills.

Preferred qualifications

Experience across the full lifecycle of high-quality video generation model training; research leadership in physical simulation, world consistency, temporal consistency or causal reasoning.
Expertise in high-quality video evaluation and human preference alignment, or efficiency optimization involving inference acceleration, VRAM compression, distillation or quantization.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

Video generationGenerative AIVideo model post-trainingInstruction dataPreference alignmentVideo quality evaluationReward modelsRLHFDPOGRPOPPOLoRAQLoRADoRAModel distillationConsistency ModelsFlow MatchingDiffusion modelsLong-video modelingCamera controlInference accelerationVRAM compressionQuantizationPhysical simulationTemporal consistencyCausal reasoningCommunicationTeamwork

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

Lead video post-training architecture
The role leads the design and implementation of video generation post-training, with emphasis on high-quality instruction data, preference alignment, and better video quality. Prepare to explain architectural decisions and how their effect on generated video was measured.
Extend video control and coherence
The job calls for long-video modeling, storyline consistency, and precise camera control. Prepare to discuss approaches to maintaining coherent content while following specific generation instructions.
Connect evaluation to alignment
The role builds video quality evaluation pipelines and multidimensional reward models, then applies alignment training using RLHF, DPO, GRPO, and PPO. Prepare to explain how evaluation signals inform training decisions.
Accelerate diffusion inference
The role researches and implements distillation and other diffusion-model inference acceleration techniques. Prepare to compare speed gains with their effect on video quality; Consistency Models and Flow Matching are named in the technical requirements.

Likely challenges to think through

Quality gains may not hold across longer videos
A likely challenge is improving visual fidelity without losing storyline or temporal consistency as video length increases. Prepare to describe an evaluation design that checks both frame-level quality and coherence across a sequence; this is a preparation hypothesis, not a reported team problem.
Reward signals may miss visible defects
A multidimensional reward model could favor measured qualities while overlooking artifacts or instruction-following failures. Prepare to explain how human preference judgments and targeted quality checks could test whether alignment gains are genuine.
Faster inference may reduce control or fidelity
Distillation and other acceleration methods can create a trade-off between generation speed and output quality, including camera-control precision. Prepare to explain how you would compare accelerated and baseline models using both efficiency and video-quality measures.

Practice your answers

Technical practice

How would you design a video-generation post-training architecture?

What to coverThe posting calls for leading post-training architecture and implementation to improve video quality. Suggested preparation: explain how you would connect instruction data, fine-tuning, preference alignment and evaluation, including how you would identify which stage caused a quality regression.

What makes instruction data useful for high-quality video generation?

What to coverThe role emphasizes high-quality instruction data. Suggested preparation: discuss coverage, instruction-to-video correspondence and data-quality checks, and explain how you would test whether data changes improve generated videos.

How would you choose among LoRA, QLoRA and DoRA for post-training?

What to coverThe requirements name all three fine-tuning approaches. Suggested preparation: compare their practical trade-offs in adaptation quality and resource use, and describe what evidence would guide a choice for a video model.

How would you improve long-video modeling and storyline consistency?

What to coverThe posting explicitly calls for both capabilities. Suggested preparation: explain how you would assess continuity across a long sequence, diagnose where it breaks down and weigh local visual quality against story-level coherence.

How would you evaluate precise camera control?

What to coverPrecise camera control is a stated capability goal. Suggested preparation: describe tests that distinguish following a requested camera movement from producing an attractive video, including how you would inspect failures over time.

How would you build a video-quality evaluation pipeline?

What to coverBuilding evaluation pipelines and using quality metrics are explicit parts of the role. Suggested preparation: explain how you would combine repeatable metrics with human assessment, check that results reflect perceived quality and compare model changes fairly.

How would you design a multidimensional video reward model?

What to coverThe role calls for multidimensional reward models and practical work with video reward models. Suggested preparation: discuss distinct quality dimensions, how you would validate agreement with human preferences and how you would detect improvements to a score that do not improve videos.

When would you use RLHF, DPO, GRPO or PPO for alignment?

What to coverThe posting names all four alignment methods. Suggested preparation: compare their training signals and implementation trade-offs, then explain how evaluation results would inform your choice rather than treating one method as universally best.

How would you accelerate diffusion-model inference through distillation?

What to coverInference acceleration is a stated responsibility, and the requirements name Consistency Models and Flow Matching. Suggested preparation: explain how you would investigate these approaches and measure speed gains against visual quality and temporal consistency.

How would you assess efficiency changes beyond inference speed?

What to coverThe preferred qualifications mention VRAM compression and quantization alongside acceleration and distillation. Suggested preparation: describe how you would measure memory use, speed and video-quality regressions before deciding whether an optimization is worthwhile.

Behavioural practice

Tell me about a time you led a research architecture decision.

What to coverThe posting calls for leadership in post-training architecture and implementation. Suggested preparation: use a real situation, your decision and actions, and the result; explain how you handled evidence that challenged your initial design.

Describe a disagreement about how to evaluate model quality.

What to coverEvaluation and teamwork are both stated priorities. Suggested preparation: give a real situation, how you worked through differing views on metrics or human assessment, and the resulting decision or lesson.

Tell me about communicating a complex research trade-off to collaborators.

What to coverThe requirements emphasize communication, while the work spans quality, alignment and efficiency. Suggested preparation: describe the audience and situation, how you made the trade-off clear, and what action or understanding resulted.

Describe a time a model improvement failed to hold up under evaluation.

What to coverThe role combines model development with quality evaluation. Suggested preparation: use a real setback, explain how you investigated it and worked with others if relevant, and state the result without overstating success.

Tell me about balancing quality against inference efficiency.

What to coverThe posting seeks both video-quality enhancement and diffusion inference acceleration. Suggested preparation: describe a real prioritization situation, the evidence and actions behind your choice, and its measurable or observable result.

Sharpen your knowledge

Topics to prepare

Video post-training architecture
Posting: The role leads architecture and implementation of video generation post-training, emphasizing instruction data and video quality. Suggested exercise: Diagram a post-training workflow from instruction-data selection through fine-tuning, alignment and evaluation; explain how each stage could affect visual quality.
Preference-alignment methods
Posting: The role calls for RLHF, DPO, GRPO and PPO, with a theoretical and practical foundation in preference alignment. Suggested exercise: Compare their training signals, data needs and trade-offs in a one-page table, then explain which questions you would test experimentally for video generation.
Video evaluation and reward modeling
Posting: Building video-quality evaluation pipelines and multidimensional Video Reward Models is a responsibility; experience with quality metrics is required. Suggested exercise: Define separate measures for visual fidelity, temporal consistency and instruction adherence, and explain where a combined reward might obscure failures.
Long-video consistency and control
Posting: The role explores long-video modeling, storyline consistency and precise camera control; world and temporal consistency are plus points. Suggested exercise: Analyze a hypothetical multi-shot prompt and list likely continuity and camera-control failure modes, distinguishing required capabilities from plus-point research areas.
Diffusion inference acceleration
Posting: The role researches distillation for diffusion inference acceleration and names Consistency Models and Flow Matching; VRAM compression and quantization are plus points. Suggested exercise: Explain how you would compare a distilled model with its baseline on video quality, inference speed and memory use.

Skills to revise

Prepare instruction data and a fine-tuning plan
Posting: High-quality instruction data and fine-tuning with LoRA, QLoRA or DoRA are explicit. Suggested exercise: Draft a small set of video instructions covering motion, storyline and camera direction; specify quality checks and a comparison plan for the named fine-tuning methods.
Build a video-evaluation prototype
Posting: The job calls for evaluation pipelines, quality metrics and multidimensional reward models. Suggested exercise: Create an evaluation specification for a small set of generated clips: record per-dimension scores, aggregate them, and inspect cases where aggregate scores disagree with the underlying dimensions.
Design an alignment experiment
Posting: Executing alignment training with RLHF, DPO, GRPO and PPO is part of the role. Suggested exercise: Write an experiment plan using video preference comparisons: identify the preference signal, choose a named alignment method, define a baseline and specify checks for quality regressions.
Plan and assess distillation
Posting: The role implements distillation to accelerate diffusion-model inference. Suggested exercise: Outline a teacher–student distillation experiment using a named approach from the posting, then define how you would measure speed gains against visual and temporal-quality changes.
Present research trade-offs clearly
Posting: Excellent communication and teamwork are required alongside research and implementation responsibilities. Suggested exercise: Prepare a brief technical explanation of one proposed post-training change, including its evidence, evaluation results you would seek and trade-offs a collaborating team would need to decide.

Your preparation checklist

Prepare a post-training project walkthrough
The posting asks for proven video generation post-training work. Select a relevant project, if available, and document your role, instruction-data choices, fine-tuning approach, results, limitations, and collaboration.
Sketch a post-training architecture
As an interview exercise, draw a pipeline from high-quality instruction data through preference alignment to video-quality evaluation. Explain design trade-offs and how you would lead implementation; this is a preparation exercise, not a stated interview requirement.
Design a video evaluation plan
The role calls for evaluation pipelines and multidimensional reward models. Prepare a sample scorecard covering visual quality, temporal and storyline consistency, and camera-control accuracy; describe metrics, human judgments, and likely failure cases.
Compare alignment and fine-tuning choices
Make a one-page comparison of RLHF, DPO, GRPO, and PPO, alongside LoRA, QLoRA, and DoRA. Be ready to explain which you have used, what evidence you can show, and where you would need to experiment.
Outline a long-video capability experiment
Propose a bounded test for storyline consistency or precise camera control, both named in the posting. Specify prompts, comparisons, evaluation criteria, and how you would diagnose failures.
Prepare an inference-acceleration trade-off analysis
The posting calls for diffusion-model inference acceleration and names Consistency Models and Flow Matching. Outline an experiment comparing quality, latency, and memory use, distinguishing measured experience from proposed work.
Write targeted questions for the team
Ask how the Sea AI team prioritizes video quality, long-video capabilities, alignment, and inference speed for social and e-commerce applications, and what evidence would define success in this role.

Availability last checked: 4 Nov 2026. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in Singapore. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…