LongE2V¶

Development line: project:longe2v · thread longe2v-development
Last event: - · 0 dated since - · Researched: 2026-09-06 · confidence: medium

What it is¶

LongE2V is an open-source research stack for event vision, fine-tuned as a LoRA on THUDM/CogVideoX-5B-I2V.

• Video reconstruction from sparse event streams.
• Long-horizon prediction from sparse event streams.
• Frame interpolation from sparse event streams.

Standard run requires Linux, NVIDIA/CUDA 12.8, Python 3.10, and event voxels; no hosted inference provider exists. The stack fits reproduction and evaluation of event-camera tasks, not standard video generation from photos.

Development line¶

• The dated line is not written up yet; what is known stands in the sections below.

What changed¶

2026-07-12 — LongE2V appeared as a SIGGRAPH 2026 project with code for a unified diffusion approach to event-based reconstruction, prediction, and frame interpolation. 2026-07-14 — The associated LongE2V model and dataset on Hugging Face were marked updated; the public path now includes ready LoRA weights and preprocessed data for local reproduction.

How to use this¶

1. Clone the repository, create a Python 3.10 environment, and install PyTorch 2.8.0 for CUDA 12.8 and project dependencies. — https://github.com/cdfan0627/LongE2V
2. Download the published LoRA file into weights/ and fetch the preprocessed data, or prepare event voxels manually. — https://github.com/cdfan0627/LongE2V
3. Choose reconstruction, prediction, or interpolation, and run the matching infer_*.sh with the DATASET value set. — https://github.com/cdfan0627/LongE2V
4. Process output frames before comparing metrics: resize, apply brightness correction for reconstruction and prediction, then run the matching metrics script. — https://github.com/cdfan0627/LongE2V

Best practices¶

• LongE2V: do not substitute standard input images for event streams; the methods condition on event voxels and solve event-camera tasks. — https://arxiv.org/abs/2607.08770
• LongE2V: use the published data structure for reproducible results; missing or extra voxel files alter event chunks, and reverse voxels are needed only for interpolation. — https://github.com/cdfan0627/LongE2V
• LongE2V: run on one GPU and preserve compatibility with CUDA 12.8 and pinned packages; multi-GPU remains untested, and smaller ATTENTION_SCORE_CHUNK_SIZE saves memory at the cost of speed. — https://github.com/cdfan0627/LongE2V
• LongE2V: do not compare raw frames against published metrics; reconstruction and prediction metrics require a separate EVREAL environment because pyiqa conflicts with core dependencies. — https://github.com/cdfan0627/LongE2V

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• LongE2V — no public changelog, tag, or GitHub Release exists, so exact dates of initial code and weight publication and subsequent semantic changes cannot be reliably recovered.
• LongE2V — availability and compatibility with arbitrary custom event data and GPUs were not confirmed by a live run; documentation confirms only the published research path.

Sources¶

Agent brief¶

• Subject: project:longe2v, thread longe2v-development, 0 dated events - → -.
• Practical note: See the sourced usage and practice sections above, including their limits.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.