๐ŸŽฌVideoModelWeekly
โ† Back to archive

Issue #11 โ€” Text-to-4D Breakthroughs & Editing Benchmarks

Published 2026-02-24 ยท 5 papers ยท 3 generation ยท 2 editing

generationDreamVideo4D ยท 2.8B

DreamVideo4D: Text-to-4D Dynamic Scene Generation

First open-source model generating dynamic 4D scenes from text prompts with consistent geometry across time.

MetricScoreฮ”
4D-FID42.3-8.7
Geometry Consistency91.2+3.4
CLIPSIM0.291+0.012
๐Ÿ“„ arXiv๐Ÿ”— GitHubDataset: Objaverse + WebVid-2M
editingFastEdit ยท 1.2B

FastEdit: Real-Time Video Editing with Latent Caching

Achieves near-real-time video editing by caching latent features and only recomputing edited regions.

MetricScoreฮ”
Edit Speed (fps)18.4+12.1
Edit Accuracy88.7%+1.2%
FVD345-22
generationLumaGen-2 ยท 6.1B

LumaGen-2: Cascaded Video Diffusion at Scale

Two-stage cascaded diffusion pipeline generating 720p videos with improved temporal coherence.

MetricScoreฮ”
VBench Total82.1+1.5
FVD (UCF-101)245-19
Aesthetic Score6.8+0.3
๐Ÿ“„ arXiv๐Ÿค— HuggingFaceDataset: LumaVid-Internal
generationControlVideo-2 ยท 2.1B

ControlVideo-2: Multi-Condition Controllable Video Generation

Supports simultaneous depth, pose, and edge control signals for precise video generation.

MetricScoreฮ”
Control Accuracy93.1%+4.2%
FVD278-33
CLIPSIM0.301+0.009
editingVideoRecolor ยท 650M

VideoRecolor: Semantic Color Grading for Video Clips

Automatic semantic-aware color grading that can restyle entire video clips while preserving content.

MetricScoreฮ”
Color Accuracy94.5%+2.8%
Temporal Consistency96.1+1.2

Want full archive access?

Upgrade to Pro for $5/month and get every past issue plus early Tuesday delivery.

See pricing โ†’