Alessandro Ferrari

Founder @argovision | AI Professor @unipv | creator at | Chairman #Py4AI conference

Computer Vision and #AI

[Alessandro Ferrari @AI_withPapers]

🎷Layered PSD Diffusion🎷

👉OmniPSD: generates layered PSD files with transparent RGBA layers (text, foreground, background), ready to edit. Online Demo💙

✅Text-to-PSD & image-to-PSD
✅Transparency-preserving RGBA-VAE

👉Review https://t.co/5IKAgAddi7

#AIwithPapers

[Alessandro Ferrari @AI_withPapers]

🐘TTSC for 3D Generative🐘

👉SpaceControl: new SOTA training-free, test-time spatial control for 3D generation. Repo announced💙

✅ETHZ, Stanford, Technion & #NVIDIA
✅Geometry-aware, no costly tuning

👉Review https://t.co/19TqCQ2hdY

#AIwithPapers

[Alessandro Ferrari @AI_withPapers]

✌️SOTA Generative SLP✌️

👉Stable Signer: sign language production as hierarchical end-to-end generation, from text to Pose2Vid. Repo with data💙

✅SLP Mixture-of-Experts
✅+48.6% over SOTA

👉Review https://t.co/XpHJj7gkLQ

#AIwithPapers

[Alessandro Ferrari @AI_withPapers]

🦄 Native Unified Multimodal 🦄

👉META's TUNA: native unified multimodal model, one continuous visual space (VAE + representation encoder). Code under review💙

✅Image/video understanding & generation
✅SOTA

👉Review https://t.co/rI6x9CGBFt

#AIwithPapers

[Alessandro Ferrari @AI_withPapers]

🥭3D Point Motion Editing🥭

👉Edit-by-Track: video motion editing via 3D point tracks: control camera & objects, remove objects, transfer motion. No code💙

✅3D track conditioner
✅Two-stage training

👉Review https://t.co/ujLbxhT4Ia

#AIwithPapers

[Alessandro Ferrari @AI_withPapers]

🎊 Pixel-Level 4D Humans 🎊

👉Dense4D (Imperial College): dense 4D correspondence of body, clothing & hair from monocular RGB, no template. Code announced💙

✅First feed-forward model for it
✅Much higher tracking accuracy

👉Review https://t.co/KpFgOIB9v5

#AIwithPapers

[Alessandro Ferrari @AI_withPapers]

🧊SOTA category-level 9D Pose🧊

👉SchemaPose (SJTU at #NeurIPS 2026): new SOTA among RGB-based methods for category-level 9D pose. Code announced💙

✅Schema-guided category structure
✅Schema-driven data generation

👉Review https://t.co/QZzU55UsHH

#AIwithPapers

[Alessandro Ferrari @AI_withPapers]

🐇Segment where you look🐇

👉KAUST (+#Adobe) rethinks how a VLM tells the segmentation decoder WHERE its referent is. Code announced💙

✅ATTNSEG replaces the [SEG] embedding
✅Attention bridges language & segmenter

👉Review https://t.co/W33JvTqjSE

#AIwithPapers

[Alessandro Ferrari @AI_withPapers]

🌵Instance-Level Video Generation🌵

👉InstanceV: the first video generation framework built for instance-level control at the architectural level. Code & Data announced💙

✅Instance-aware masked cross-attention
✅New benchmark

👉Review https://t.co/ZIxCTjWDgd

#AIwithPapers

[Alessandro Ferrari @AI_withPapers]

🕶️ Seeing without Pixels 🕶️

👉Can you understand a video without its pixels, just from the camera trajectory? Deepmind (+UTexas) investigates it💙

✅Contextualized trajectory encoding
✅Beats heavy vision models

👉Review https://t.co/B9X33v4JHs

#AIwithPapers

[Alessandro Ferrari @AI_withPapers]

🔥 Smell Like Vision Spirit 🔥

👉New York Smells: in-the-wild dataset of paired vision & smell for cross-modal learning. With the lights out, it's less dangerous. Data💙

✅Columbia, Cornell & Osmo Labs
✅7,000 smell-image pairs

👉Review https://t.co/mrjypNuC90

#AIwithPapers

[Alessandro Ferrari @AI_withPapers]

🍓MotionV2V: Editing Motion in Video🍓

👉Google's MotionV2V: edit the motion of a video by controlling its change via diffusion. Impressive. Repo TBA💙

✅Motion counterfactual video pairs
✅Editing via user motion trajectories

👉Review https://t.co/aNufRtxEfS

#AIwithPapers

[Alessandro Ferrari @AI_withPapers]

🌩️ Cloud4D in time 🌩️

👉Cloud4D: physically-realistic 3D cloud fields from ground cameras at 25 m and 5 s resolution. Data released, repo coming💙

✅Homography-guided 2D-3D transformer
✅10× finer temporal res

👉Review https://t.co/oPiqc206dq

#AIwithPapers

[Alessandro Ferrari @AI_withPapers]

🧪 EfficientSAM3 is out 🧪

👉lightweight SAM3 students via Progressive Hierarchical Distillation. Code coming💙

✅Flexible accuracy-latency trade-offs
✅Compact Perceiver module on SA-V

👉Review https://t.co/kIAgvmo1Zt

#AIwithPapers

[Alessandro Ferrari @AI_withPapers]

🦞Single Synthetic Image per Class🦞

👉MIT's Linear Gradient Matching (H/T Torralba): ONE synthetic image per class to train linear classifiers. Repo available💙

✅Beats all real-image baselines
✅Strong out-of-distribution

👉Review https://t.co/EV2ECLuhAW

#AIwithPapers

[Alessandro Ferrari @AI_withPapers]

🍕 Upsample Anything 🍕

👉Upsample Anything: universal, training-free up-sampler via lightweight test-time optimization. No code, but relevant💙

✅≈0.419s per 224×224 image🔥
✅SOTA on segmentation & depth

👉Review https://t.co/MGg4IN5fAi

#AIwithPapers

[Alessandro Ferrari @AI_withPapers]

🍯Unwrapping of 3D Meshes🍯

👉PartUV: part-based UV unwrapping for 3D meshes, combining learned part priors with geometric cues. Repo released💙

✅Compact part-aligned charts
✅Fewer fragments, same low distortion

👉Review https://t.co/5MtlAgmPLj

#AIwithPapers

[Alessandro Ferrari @AI_withPapers]

🔥 SAM 3/3D 🔥

👉#META released SAM 3: detection, segmentation & tracking in images/video via text, exemplar & visual prompts💙

✅Doubles accuracy of existing systems
✅SAM 3D: open models & data for 3D

👉Review https://t.co/cutsKUD0Sc

#AIwithPapers

[Alessandro Ferrari @AI_withPapers]

⌚ Multi-Shot Video Segmentation ⌚

👉Fudan tackles multi-shot video object segmentation (MVOS). Benchmark & repo under Apache 2.0💙

✅SAAS: first semi-supervised MVOS method
✅Transition Mimicking Augmentation

👉Review https://t.co/29sVC2MFpt

#AIwithPapers

[Alessandro Ferrari @AI_withPapers]

🌩️ It's "Time-to-Move" 🌩️

👉Technion + Nvidia Time-to-Move: training-free motion & appearance control for I2V diffusion (Wan 2.2, CogVideoX, SVD). Impressive!

✅Control via crude animations
✅Dual-clock denoising

👉Review https://t.co/89E7xUrADY

#AIwithPapers