Baolu Li

Ph.D Candidate, Dalian University of Technology

[Odyssey @odysseyml]

Today we're launching Odyssey-3, the most powerful foundation world model yet.

It sets a new state of the art on Physics-IQ, and powers robots, trains AIs, and generates interactive experiences. It's really cool.

Experience the model today, all for free!

[Ziwei Liu @liuziwei7]

🛠️Tool-Centric Reasoning in Egocentric Videos🛠️

🧠 Tool use separates us from other animals.

But can AI understand why we reach for a tool, which tool we choose, and how we use it to get things done?

@ropedia_ai builds #EgoTools to find out.

- Project: https://t.co/2e64e2l2V6

[AK @_akhaliq]

EgoTools

Towards Tool-Centric Reasoning in Real-World Egocentric Videos

paper: https://t.co/DYdf8dBPJg

[Baolu Li @calm_lu_312]

Our new work on interactive video world models: Oneira 🌍

Explore generated worlds, interact with newly discovered objects, and make changes that persist as you keep exploring.

📄 Paper: https://t.co/8v4nCXnsnM
🎬 Demos: https://t.co/aFTGB50YGr

[Xindi Yang @Madaoer_Yxd]

🌍 Introducing Oneira, an interactive video world model where a coding agent keeps an explicit world state.

Objects that appear as you explore become interactive, and your changes persist over long horizons.

Project: https://t.co/DhvJxnZb2j
Paper: https://t.co/7e4uo2Iqu2

[HU Wenbo @wbhu_cuhk]

Introducing GAE🥳🥳

Generate once. Decode RGB + point clouds.

GAE generates directly in a geometry-native latent space. The same generated latent supports both appearance and 3D geometry.

The key choice is where generation happens. GAE compresses geometry-foundation features into a shared latent space. A conditional flow model generates directly in that space; RGB and geometry are decoded from its final state.

The video pairs show camera-conditioned world generation; the character pairs show text-to-image generation. These native showcase clouds are not reconstructed from generated RGB.

Code: https://t.co/euE8mZKI4G
Page: https://t.co/M0dMYwIOPp

[🍯 @gpumaxxer]

fake life maxxing 💀

with Higgsfield Genjutsu

��: @iinsightri in IG

[Alexandre Morgand @Almorgand]

"AgentSTAR: Agentic Shape Tracking and Reconstruction from Monocular Videos"

VLM agent+optimization to reconstruct an object geometry+kinematic structure from RGB video. Pose+articulation through time including occlusions, transparent objects. No dense pixel correspondences.

[Jianglong Ye @jianglong_ye]

Evolving dexterity with GPT-6 Astra 🖐️

Been trying Astra recently. Its zero-shot dexterous manipulation is already quite surprising. More interesting is seeing it learn and improve through simulation training, from pen spinning and Rubik's Cube to hammer use.

The real goal would be to evolve this dexterity in the real world.

https://t.co/GZjQTsklYc

[AK @_akhaliq]

LongHorizon-Harness

Advancing Long-Horizon Agents for Real-World Tasks

paper: https://t.co/ts89ezenct

[DailyPapers @HuggingPapers]

NVIDIA just released Audex on Hugging Face

A unified audio-text MoE that handles speech, sound, and reasoning

without regressing on text intelligence.

30B params. 3B active. 1M context.

[Matthias Niessner @MattNiessner]

📢TriFlow: Generating Artist-Like 3D Mesh Topology via Nearest-Vertex Vector Fields (ECCV'26)📢

Compact 3D meshes with clean, artist-like triangle topology - structured connectivity you'd expect a human modeler to make🙂

🌐https://t.co/PsrCkwUrCJ
▶️https://t.co/nlLTliWEEa

[Kaede Shiohara @KaedeShioharaCS]

We present GUSH3R🫟, a feed-forward approach for dynamic human-scene 3DGS reconstruction.

Our model generates per-pixel scene gaussians and per-vertex dynamic human gaussians by dual decoders built on top of Human3R⚡️

Project page (with arXiv and code):
https://t.co/ATOiijnZRu

[Reid Hannaford @reidhannaford]

AI filmmaking is insane!

House of the Dragon reportedly costs $20 million per episode.

Now you can create Hollywood-level dragon shots from your bedroom with Seedance.

My workflow:

1. Make a start frame in Midjourney

2. Block-out in Blender, animate the camera

3. Feed both to Seedance

You can build a whole cinematic universe from a handful of blocks.

[Richard Palethorpe @jichiep]

depth-anything.cpp combined with a phone camera and some SLAM techniques can create voxelated worlds from time-lapse videos.

I'm not even sure a phone camera is in-distribution as a lot of the depth information is low confidence. All of the far objects below would be filtered if we demanded confidence, but we still get identifiable buildings in the background.

[AK @_akhaliq]

World Tracing

Generative Pixel-Aligned Geometry Beyond the Visible

[DailyPapers @HuggingPapers]

Real-time video-to-video editing on consumer hardware

SANA-Streaming runs 1280×704 at 24 FPS end-to-end on one RTX 5090, with its DiT core hitting 58 FPS. A hybrid diffusion transformer paired with cycle-reverse regularization keeps edits temporally coherent without needing long paired videos.

[Bohan Sun 🟥🟩🟦 @sbh_studio]

tried video spatial reframing, captured by iphone 14pro
#spatialreframing #4DGS

[Pierrick Chevallier | IA @CharaspowerAI]

🚨PromptShare🚨

🎥 One simple prompt to generate a full grid of camera angles using Banana Pro
Perfect to explore a character or scene from every perspective 👇

Prompt:
Create a 3x3 cinematic grid featuring the same characters and scene, each frame with a different cinematic camera angle and shot type.
Each panel represents a unique cinematic angle

[Yaseen Khan Gul @YaseenK7212]

Hero shot of a friend chicken thigh🤯 On Nano Banana 3.0

Prompt:

{
"food_photography_series": {
"subject": "Fried Chicken Thigh",
"technical_specs": {
"style": "Hyper-realistic hero shot",
"aspect_ratio": "3:4",
"lighting": "Cinematic / Studio"
},
"variations": [
{
"theme": "Tartar Refresh",
"sauce": "Thick tartar sauce",
"accents": ["capers", "dill", "chopped pickles"],
"background": "Bright lemon-yellow",
"prompt": "Hyper-realistic hero shot of a fried chicken thigh in midair with thick tartar sauce bursting around it, flecks of capers, dill, and chopped pickles frozen in motion. The sauce flows like creamy waves, dancing around the crispy surface. Bright lemon-yellow background with moody studio backlight creates a gourmet, refreshing tone ultra-detailed commercial look, --ar 3:4"
},
{
"theme": "White Truffle Luxury",
"sauce": "Creamy white truffle sauce",
"accents": ["shaved mushrooms", "fine porcini powder"],
"background": "Warm yellow-beige with soft haze",
"prompt": "Hyper-realistic hero shot of a luxurious fried chicken thigh floating in the center, surrounded by a splash of creamy white truffle sauce, with shaved mushrooms and fine powder of porcini dusting the motion. Golden skin shines under cinematic key light, background in warm yellow-beige hues with soft haze premium gourmet commercial, --ar 3:4"
},
{
"theme": "Smoked Cheese Indulgence",
"sauce": "Molten smoked cheese fondue",
"accents": ["golden crouton crumble"],
"background": "Creamy yellow with taupe haze",
"prompt": "Hyper-realistic hero shot of a crisp fried chicken thigh surrounded by a molten splash of smoked cheese fondue mid-air, with bits of golden crouton crumble suspended in orbit. Deep lighting creates dimensional drama on the crispy crust. The backdrop blends creamy yellow with taupe haze indulgent cheese-forward visual, --ar 3:4"
},
{
"theme": "Miso Umami",
"sauce": "Miso butter glaze and umami cream",
"accents": ["toasted seaweed powder", "sesame seeds"],
"background": "Muted beige",
"prompt": "Hyper-realistic hero shot of a rich fried chicken thigh hovering mid-air, glazed with miso butter and surrounded by a dynamic splash of umami cream, flecked with toasted seaweed powder and sesame seeds. Dramatic soft shadows sculpt the skin texture. Muted beige background, clean and warm umami-focused food ad, --ar 3:4"
}
]
}
}

[David Roberts @recap_david]

Kling 2.5 Turbo + Nano Banana Pro

Perfect for food-based ads.
Get this video template for free in Fabric AI (link below)

Image Prompt 1:

[Overhead, top-down food photography of a vibrant, healthy steak salad bowl on a light beige stone surface. A shallow ceramic bowl filled with finely chopped curly kale as the base. Medium-rare sliced steak arranged neatly across the center, pink interior with seared edges and visible seasoning. On one side, fanned avocado slices with cracked black pepper. Bright red diced bell peppers, crumbled white cheese (feta-style), and finely chopped greens distributed evenly around the bowl. A lemon wedge tucked against the edge of the bowl. A blue-handled fork resting inside the bowl, angled slightly toward the center. Natural soft daylight from above, minimal shadows, clean editorial food styling. Sharp focus, high detail, realistic textures, fresh and appetizing. Modern cookbook / Instagram food photography aesthetic. No hands, no text, no branding, no clutter.]

Image Prompt 2:

[Using the original steak salad image as the sole ingredient reference, create a clean, vertically stacked exploded-view visualization of the same salad.
The ingredients are arranged in a strict bottom-to-top order, evenly spaced along a single centered vertical axis, with symmetry, alignment, and visual balance.
Bottom layer (base of the salad):
Finely chopped curly kale and mixed leafy greens, forming a soft, natural pile that anchors the composition.
Middle layers (core ingredients), stacked upward in this order:
– Diced red bell peppers and chopped greens, evenly distributed
– Crumbled white cheese (feta-style), centered and proportionally scaled
– Medium-rare sliced steak, laid flat and neatly fanned, pink interior visible
– Avocado slices, evenly cut and symmetrically fanned
Top layer (finishing elements):
Lemon wedge and light vinaigrette droplets, placed delicately at the top of the image to signal freshness and completion.
All ingredient layers are parallel, evenly spaced, and centered, with no rotation, tilt, or perspective distortion.
Ingredients appear to float gently while maintaining realistic proportions, textures, and color accuracy.
Add short, minimal annotations with thin leader lines.
Alternate caption placement from left to right as you move up the stack to create visual balance and avoid crowding.
Example annotations:
– Leafy greens: “Fresh base, rich in nutrients”
– Steak: “High-quality protein”
– Avocado: “Healthy fats, creamy balance”
– Red pepper: “Natural sweetness & antioxidants”
– Lemon & vinaigrette: “Light finish enhancing natural flavors”
Background is pure white or very light neutral, matte and distraction-free.
Lighting is soft, even, and shadow-minimized with a clean editorial food-photography feel.
Style is premium food photography combined with a technical exploded diagram, suitable for marketing, nutrition education, or app UI.
No hands, no bowl, no clutter, no branding, no dramatic shadows.]

Kling 2.5 Turbo Prompt:

A high-angle, cinematic studio shot of a fresh steak salad in a white ceramic bowl, centered on a clean white background. The salad is fully assembled: finely chopped curly kale and leafy greens, medium-rare sliced steak, avocado slices, diced red bell peppers, crumbled white cheese, and a lemon wedge.
After a brief moment of stillness, the salad bursts upward in a controlled, elegant explosion, with each ingredient separating cleanly and moving upward along a vertical axis. The motion is smooth, slow, and weightless—no chaos, no spinning—creating a precise, visually satisfying deconstruction.
The ingredients settle into a perfectly aligned exploded-view composition, hovering in mid-air in distinct horizontal layers, evenly spaced and centered:
– Leafy greens at the bottom
– Red bell peppers and crumbled cheese above
– Medium-rare sliced steak laid flat and neatly fanned
– Avocado slices symmetrically arranged
– Lemon wedge and light vinaigrette droplets at the top
Once the ingredients are fully separated and stable, minimal technical annotation lines and labels fade in, alternating left and right for balance. Text is clean, modern, and readable, connected with thin leader lines, never overlapping the ingredients.
Lighting is soft and diffuse, studio-style, with minimal shadows.
Motion is slow-motion and cinematic, emphasizing clarity and elegance.
Ultra-realistic food textures, high detail, premium editorial aesthetic.
Camera remains locked and steady throughout.
No hands, no people, no clutter, no branding, no background movement.
End on the fully exploded, clearly labeled ingredient view.]