I joined Agile Robots SE as a research intern working on world models for robot dexterous manipulation.
Wenbo Ji 嵇文博
I am an M.Sc. student at the Technical University of Munich and a research intern at Agile Robots SE. My work focuses on video generation and ML systems for video diffusion, building on a research background in 3D/4D representation and video tracking.
I am primarily seeking industry roles as a Research Engineer, Research Scientist, or Member of Technical Staff (MTS) in video generation, multimodal AI, embodied AI, and spatial intelligence. I am also open to Ph.D. opportunities in these areas.
Background
Previously, I collaborated with Prof. Daniel Cremers, Prof. Nassir Navab, and Prof. Benjamin Busam on 3D reconstruction and tracking. I earned an M.Sc. in Computer Science from Tongji University and a B.Sc. in Information and Computing Science from Nanjing Tech University.
News
Our paper CSG-Fusion received the Best Paper Award at the ICCV 2025 Workshop E2E3D.
I graduated from Tongji University with a Master's degree in Computer Science.
More news
Our paper LiteTracker is accepted by MICCAI 2025.
Our paper RE0 is accepted by ICRA 2025.
I joined ImFusion as a research intern on dense video point tracking.
I joined TUM-CVG and TUM-DI-LAB under the co-supervision of Yan Xia and Chuanxia Zheng.
I joined the Technical University of Munich as a Master's student in Electrical Engineering and Information Technology.
Research Interests
-
Controllable video generation
Camera and motion conditioning for video generation, with a focus on expressive human motion, identity consistency, and scene interactions.
-
Multimodal AI infrastructure
Large-scale multi-GPU training and efficient, reliable inference for multimodal models, spanning distributed execution, GPU kernel optimization, and correctness in caching and serving.
-
Spatial intelligence and embodied AI
3D/4D representation and video tracking for dynamic scene understanding, alongside video world models and cross-embodiment generation for robot learning.
Selected Publications
* equal contribution, † corresponding author
CSG-Fusion: Consistent Sparse-View Gaussian Splatting via Matching-based Fusion
LiteTracker: Leveraging Temporal Causality for Accurate Low-latency Tissue Tracking
Open Source Contributions
An inference and serving framework for large language, vision-language, and diffusion models.
Contributions: Native diffusion model integration, sequence-parallel numerical consistency, and CFG-aware cache lifecycles.
Selected PRs
Native diffusion execution
-
Native Ovis-Image 7B execution with checkpoint mapping, CFG, and TP/SP. Single-B200 exact-tier output matches pinned Diffusers bit-for-bit; four-B200 replay is validated within BF16 control error.
Sharded attention consistency
-
Pad exact-tier SDPA queries to align TP/Ulysses head-shard reduction paths. Operator and two-B200 checks are bit-identical; whole-model differences remain. Default lossless-tier behavior is unchanged.
Request-local cache state
-
Isolate TeaCache CFG branches and reset cached steps and residuals per request. Reused Wan instances match fresh execution bit-for-bit on 1/2 B200s under serial/parallel CFG.
A modular diffusion library for image, video, and audio generation.
Contributions: CFG-aware cache integration and request-lifecycle validation for Wan video-to-video pipelines.
Selected PRs
Branch-aware video caching
-
Separate Wan V2V CFG cache contexts for TaylorSeer, FirstBlockCache, and MagCache. One-B200 Wan1.3B smoke tests verify cache-off parity and request reset; speed and generation quality are unqualified.
A diffusion model engine for image, video, and audio generation, with support for inference and training.
Contributions: Memory-efficient tiled Qwen-Image VAE encoding and decoding for high-resolution generation.
Selected PRs
Memory-efficient VAE execution
-
Blend overlapping QwenImageVAE tiles for memory-efficient encoding and decoding. One B200, 2048² BF16, tile/stride 128/64: isolated encode/decode peaks 7.77/8.53 → 2.15/2.43 GiB; tiled outputs differ slightly.
A serving framework for multimodal generation across text, images, audio, and video.
Contributions: CFG-aware TeaCache isolation and request-failure recovery for diffusion serving.
Selected PRs
Guidance-aware cache ownership
-
Separate conditional/unconditional TeaCache state through shared CFG contexts. Regressions cover branch-local updates and fallback without an initialized distributed group.
Request failure recovery
-
Isolate failed requests and clean up partially submitted batches. GPU regressions preserve engine reuse for later requests, covering list and generator inputs.
A distributed inference engine that scales diffusion transformers across multiple GPUs.
Contributions: Wan tensor-layout optimization, sequence-parallel numerical consistency, and distributed VAE coordination.
Selected PRs
Contiguous patch-embedding layouts
-
Restore contiguous Wan patch embeddings for cuDNN execution. Wan1.3B BF16, 480×832, 81 frames, 50 steps: two-B200 Ulysses/Ring median latency −6.3%/−7.0%, with bit-identical outputs.
Valid-token attention boundaries
-
Propagate valid-token bounds through padded sequence shards and joint attention, guarding unsupported layouts. Eight-B200-node tests match selected ZImage/Cosmos3/QwenImage/LingBot references exactly; FLUX2 retains a small replay error.
Distributed execution ownership
-
Give prompt sharding, output gathering, and parallel VAE decoding explicit ownership. An eight-B200 DP/Ulysses/CFG matrix verifies missing/duplicated-output fixes and coordinated VAE collectives.
GPU attention kernels that accelerate exact attention while reducing memory use.
Contributions: SM100 attention scheduling, asymmetric backward correctness, and NaN-safe KV-cache loads.
Selected PRs
Cache-safe attention forward
-
Sequence-bound nonpaged V-cache TMA loads and zero-fill unused rows to prevent NaN contamination. B200: 146 GPU regressions and memory checks pass; valid-value NaNs and paged KV retain their contracts.
Cache locality and launch geometry
-
Adapt co-authored CLC L2 swizzling while preserving STATIC 2CTA. B200 BF16 causal, 180 1CTA/CLC shapes: 1.039× geomean speedup; 42 shapes regress.
Asymmetric-head backward correctness
-
Separate dV TMEM/SMEM staging widths from Q/K for asymmetric head dimensions. Targeted MHA/GQA gradient regressions cover silent backward errors.
An inference and post-training framework for accelerating video generation.
Contributions: Fused Triton layout transforms, Wan-VACE model integration, and SM100 attention quantization kernels.
Selected PRs
Fused sparse-attention data movement
-
Fuse VSA64 scatter and layout transforms into one Triton kernel. Wan2.1-VSA-T2V-14B, 50 steps, 1/2 B200s: 1.155×/1.145× speedup with bitwise-equal tested outputs (SM100 BF16, D128, tile64; inference only).
Sequence-parallel control streams
-
Shard Wan-VACE 1.3B/14B image, video, and mask control streams across sequence-parallel ranks. Teacher-forced checks compare against Diffusers; FP32 SP2 uses 1e-5 tolerance. BF16 parallel outputs are not bitwise-equivalent.
Long-sequence attention tiling
-
Optimize Attn-QAT full-tile/masked-tail loops and 64×128 backward tiling. Fixed-batch Wan2.1-1.3B SP2, two B200s: forward/backward/AdamW step latency falls 10.9%, preserving tested outputs and gradients bit-for-bit.
RL post-training infrastructure with deterministic operators for consistent computation across training and rollout engines.
Contributions: Deterministic attention and memory-efficient log-probability kernels, plus TP/EP rollout–recompute parity.
Selected PRs
Batch-invariant attention kernels
-
Fixed-order Triton H3 attention without atomics or split-KV. One H100, BF16 non-causal forward/backward: batch/head invariant, up to 48× vs strict deterministic reference at 8,349 tokens; still 1.6–2.3× slower than SDPA.
Rollout/recompute reproducibility
TP4 validation → EP4 extension
-
Map Qwen3-Next TP4 weights across vLLM rollout and Megatron recompute. Four B200s, EP1: 64 replay cells validate bitwise parity through two updates and checkpoint reload.
-
Extend TP4 replay to EP4 with fixed-order expert routing. One four-B200 node: 16/16 rollout/recompute and 97/97 batch-invariance cases pass; EP4 training and checkpoint export stay disabled.
Bounded-memory log-probability backward
-
Fuse direct-write logp gradients with a bounded chunked fallback. One B200, BF16 32768×151936 logits: peak memory 46.37 → 9.27 GiB including gradients; some tuned cross-entropy alternatives are faster.
Experiences
Video World Model for Robot Dexterous Manipulation
- Developing a cross-embodiment video generation method that translates egocentric human demonstrations into robot-domain videos for downstream policy learning.
Human Motion Video Diffusion
- Developing a camera-controlled video diffusion model for controllable synthesis of human motion and scene interactions across changing viewpoints.
Human Head Avatar Animation
- Led the development of ViDS, an identity-preserving video diffusion method for long-form portrait animation.
- ViDS ranked first on 8 of 13 VFHQ metrics, improving reenactment quality and identity preservation.
Dense Point Tracking
- Implemented LiteTracker’s online inference and EMA-flow initialization, and conducted low-latency tracking experiments.
- LiteTracker remained competitive on STIR and SuPer while running ~7× faster than its predecessor and 2× faster than prior state of the art.
3D Scene Decomposition
- Designed and evaluated CSG-Fusion; received Best Paper at ICCV Workshop E2E3D.
- Improved ScanNet++ PSNR by 2.8 dB over Splatt3R at 90% overlap with ~124K fewer Gaussians, and demonstrated zero-shot generalization on DTU.
Large Scale 3D Scene Reconstruction
- Contributed to the development and experimental evaluation of a large-scale 3D scene reconstruction pipeline.
Blog
Thoughts on research, 3D, video generation, and the occasional in-between.
Latest posts are available on the blog page.
Projects
LiteAvatar - WASM Version
A lightweight audio-driven 2D avatar solution that runs entirely in the browser using WASM based on Lite-avatar. No backend server required.
Gallery
Interactive Thermal Receipt Simulation
Receipt Deformation Simulation
Duck Pool: Interactive Water Simulation
Dive into an interactive water simulation where ducks float, collide, and create ripples.
Gallery