🔬 ML PAPERS Дайджест (Image,video,text - arXiv, Feb 13 2026)

🔥 MonarchRT — Efficient attention for real-time video generation via Monarch matrix factorization. Makes autoregressive video DiT viable.
→ arxiv.org/abs/2602.12271

🔥 DreamID-Omni — Unified human-centric audio-video gen. Multi-person identity + voice disentanglement in one framework.
→ arxiv.org/abs/2602.12160

UniT — Unified multimodal CoT with test-time scaling
→ arxiv.org/abs/2602.12279

UniDFlow — Discrete flow matching for multimodal understanding + generation + editing
→ arxiv.org/abs/2602.12221

DeepGen 1.0 — Lightweight unified model for image gen & editing
→ arxiv.org/abs/2602.12205

FAIL — Adversarial imitation learning for flow matching post-training (no reward model needed)
→ arxiv.org/abs/2602.12155

GigaBrain-0.5M — VLA from world model RL (robotics)
→ arxiv.org/abs/2602.12099

ладно seedream 2 победил все

(но это не точно)

Qwen AI Slides

Под капотом Qwen3 Agent и Qwen-Image 2.0

Пробуем тут:
https://chat.qwen.ai/?inputFeature=slides

Нужно зарегаться и залогиниться.

Стартапы-слайдоделы напряглись

@cgevent

☕️ Daily AI Digest — Feb 14, 2026 (Valentine's Edition 💘)

━━━━━━━━━━━━━━━━━━━━

🔬 PAPERS & RESEARCH

🔥 DeepGen 1.0 — Lightweight 5B unified model for image gen + editing. Beats 80B HunyuanImage by 28% on WISE, 27B Qwen-Image-Edit by 37%. Open-source weights, code, datasets.
📄 arxiv.org/abs/2602.12205
💻 github.com/DeepGenTeam/DeepGen

Latent Forcing — Pixel-space image generation matching latent diffusion quality without a separate decoder. Joint latent+pixel denoising with tuned noise schedules.
📄 arxiv.org/abs/2602.11401

Ctrl&Shift — Geometry-aware object manipulation in images/videos. Move & reorient objects while preserving scene realism, no explicit 3D reconstruction needed.
📄 arxiv.org/abs/2602.11440

Active-Zero — Self-evolving VLMs through active environment exploration (self-play but for vision models). Shifts from passive datasets to active visual data seeking.
📄 arxiv.org/abs/2602.11241

C-JEPA — Causal world models via object-level latent interventions. Extends JEPA from image patches to object-centric representations with counterfactual reasoning.
📄 arxiv.org/abs/2602.11389

New day, new JEPA. Extension, anyone? :D

Если не нравится сюжет, теперь можно переснять
Ждем римейков последних сезонов Игры престолов :D

твит