IA Lab

AI experiments, AI+Art crossovers, and curated external articles about artificial intelligence and creative technology.

0 5 0 Comment on Qwen-Image-2.1 Writes Its Own Alpha Channel — The Open Model That Generates Real Transparent PNGs, and Now Does It in 8 Steps
Qwen-Image-2.1 Writes Its Own Alpha Channel — The Open Model That Generates Real Transparent PNGs, and Now Does It in 8 Steps Qwen-Image-2.1 Writes Its Own Alpha Channel — The Open Model That Generates Real Transparent PNGs, and Now Does It in 8 Steps Qwen-Image-2.1 Writes Its Own Alpha Channel — The Open Model That Generates Real Transparent PNGs, and Now Does It in 8 Steps Qwen-Image-2.1 Writes Its Own Alpha Channel — The Open Model That Generates Real Transparent PNGs, and Now Does It in 8 Steps

Qwen-Image-2.1 Writes Its Own Alpha Channel — The Open Model That Generates Real Transparent PNGs, and Now Does It in 8 Steps

Hook Here is the boring step every 3D and concept artist does ten times a day: generate…

Issanou 30 October 2026
0 6 0 Comment on Framestore Just Played a Character With Gaussian Splats in Superman — 192 Cameras, 1.6 Million Splats a Frame, and a Pipeline Where Nothing Loaded Natively
Framestore Just Played a Character With Gaussian Splats in Superman — 192 Cameras, 1.6 Million Splats a Frame, and a Pipeline Where Nothing Loaded Natively Framestore Just Played a Character With Gaussian Splats in Superman — 192 Cameras, 1.6 Million Splats a Frame, and a Pipeline Where Nothing Loaded Natively Framestore Just Played a Character With Gaussian Splats in Superman — 192 Cameras, 1.6 Million Splats a Frame, and a Pipeline Where Nothing Loaded Natively Framestore Just Played a Character With Gaussian Splats in Superman — 192 Cameras, 1.6 Million Splats a Frame, and a Pipeline Where Nothing Loaded Natively

Framestore Just Played a Character With Gaussian Splats in Superman — 192 Cameras, 1.6 Million Splats a Frame, and a Pipeline Where Nothing Loaded Natively

A big-budget movie just used Gaussian splats to play a character, not a background. In Superman, the…

Issanou 30 October 2026
0 9 0 Comment on BrickBench — Stanford Handed 11 Frontier Agents a Box of LEGO and Asked Them to Build Something That Actually Stands Up
BrickBench — Stanford Handed 11 Frontier Agents a Box of LEGO and Asked Them to Build Something That Actually Stands Up BrickBench — Stanford Handed 11 Frontier Agents a Box of LEGO and Asked Them to Build Something That Actually Stands Up BrickBench — Stanford Handed 11 Frontier Agents a Box of LEGO and Asked Them to Build Something That Actually Stands Up BrickBench — Stanford Handed 11 Frontier Agents a Box of LEGO and Asked Them to Build Something That Actually Stands Up

BrickBench — Stanford Handed 11 Frontier Agents a Box of LEGO and Asked Them to Build Something That Actually Stands Up

Give an AI a prompt, get a picture. That is the old game. BrickBench plays a harder…

Issanou 28 October 2026
0 15 0 Comment on UniForge3D — Peking University’s Model Reconstructs What Your Photos Saw and Generates What They Didn’t, in One Shared Space, From a Few Unposed Shots
UniForge3D — Peking University’s Model Reconstructs What Your Photos Saw and Generates What They Didn’t, in One Shared Space, From a Few Unposed Shots UniForge3D — Peking University’s Model Reconstructs What Your Photos Saw and Generates What They Didn’t, in One Shared Space, From a Few Unposed Shots UniForge3D — Peking University’s Model Reconstructs What Your Photos Saw and Generates What They Didn’t, in One Shared Space, From a Few Unposed Shots UniForge3D — Peking University’s Model Reconstructs What Your Photos Saw and Generates What They Didn’t, in One Shared Space, From a Few Unposed Shots

UniForge3D — Peking University’s Model Reconstructs What Your Photos Saw and Generates What They Didn’t, in One Shared Space, From a Few Unposed Shots

Point a few phone photos at an object from random angles. No turntable, no camera tracking, no…

Issanou 28 October 2026
0 16 0 Comment on Hesai Kosmo — The Lidar Giant Just Built a Scanner That Feeds Gaussian Splats Straight Into World Models, and Fills the Gaps With Generative AI
Hesai Kosmo — The Lidar Giant Just Built a Scanner That Feeds Gaussian Splats Straight Into World Models, and Fills the Gaps With Generative AI Hesai Kosmo — The Lidar Giant Just Built a Scanner That Feeds Gaussian Splats Straight Into World Models, and Fills the Gaps With Generative AI Hesai Kosmo — The Lidar Giant Just Built a Scanner That Feeds Gaussian Splats Straight Into World Models, and Fills the Gaps With Generative AI Hesai Kosmo — The Lidar Giant Just Built a Scanner That Feeds Gaussian Splats Straight Into World Models, and Fills the Gaps With Generative AI

Hesai Kosmo — The Lidar Giant Just Built a Scanner That Feeds Gaussian Splats Straight Into World Models, and Fills the Gaps With Generative AI

Hesai, a lidar heavyweight, launched Kosmo on October 15, 2026: a handheld 8K scanner that turns a 15-minute room walk into aligned Gaussian splats, meshes and depth, then uses generative AI to fill the gaps and label objects for world-model training.

Issanou 25 October 2026
0 13 0 Comment on Helix World — The First Real-Time World Model You Can Hear: One Image In, an Explorable Scene Out, With Stereo Sound That Turns When You Do
Helix World — The First Real-Time World Model You Can Hear: One Image In, an Explorable Scene Out, With Stereo Sound That Turns When You Do Helix World — The First Real-Time World Model You Can Hear: One Image In, an Explorable Scene Out, With Stereo Sound That Turns When You Do Helix World — The First Real-Time World Model You Can Hear: One Image In, an Explorable Scene Out, With Stereo Sound That Turns When You Do Helix World — The First Real-Time World Model You Can Hear: One Image In, an Explorable Scene Out, With Stereo Sound That Turns When You Do

Helix World — The First Real-Time World Model You Can Hear: One Image In, an Explorable Scene Out, With Stereo Sound That Turns When You Do

Every world model we have covered could draw you a place. None of them could let you…

Issanou 25 October 2026
0 20 0 Comment on Kandinsky 6.0 Video — The Open Model That Generates the Picture AND the Synced Sound in One Pass, Lip-Sync Included, MIT-Licensed
Kandinsky 6.0 Video — The Open Model That Generates the Picture AND the Synced Sound in One Pass, Lip-Sync Included, MIT-Licensed Kandinsky 6.0 Video — The Open Model That Generates the Picture AND the Synced Sound in One Pass, Lip-Sync Included, MIT-Licensed Kandinsky 6.0 Video — The Open Model That Generates the Picture AND the Synced Sound in One Pass, Lip-Sync Included, MIT-Licensed Kandinsky 6.0 Video — The Open Model That Generates the Picture AND the Synced Sound in One Pass, Lip-Sync Included, MIT-Licensed

Kandinsky 6.0 Video — The Open Model That Generates the Picture AND the Synced Sound in One Pass, Lip-Sync Included, MIT-Licensed

Hook Most open video models hand you a silent clip and leave the sound to you. Kandinsky…

Issanou 25 October 2026
0 17 0 Comment on Meshy Edit — AI 3D Finally Gets an Undo for One Wrong Part: Select It, Describe the Fix, Keep Everything Else
Meshy Edit — AI 3D Finally Gets an Undo for One Wrong Part: Select It, Describe the Fix, Keep Everything Else Meshy Edit — AI 3D Finally Gets an Undo for One Wrong Part: Select It, Describe the Fix, Keep Everything Else Meshy Edit — AI 3D Finally Gets an Undo for One Wrong Part: Select It, Describe the Fix, Keep Everything Else Meshy Edit — AI 3D Finally Gets an Undo for One Wrong Part: Select It, Describe the Fix, Keep Everything Else

Meshy Edit — AI 3D Finally Gets an Undo for One Wrong Part: Select It, Describe the Fix, Keep Everything Else

Until now, fixing one wrong detail on an AI-generated 3D model meant regenerating the whole thing —…

Issanou 21 October 2026
0 18 0 Comment on MOSS-Image — Fudan’s OpenMOSS Team Open-Sourced a 5B Image Model You Prompt With a JSON Layout, So Text and Characters Land Exactly Where You Put Them
MOSS-Image — Fudan’s OpenMOSS Team Open-Sourced a 5B Image Model You Prompt With a JSON Layout, So Text and Characters Land Exactly Where You Put Them MOSS-Image — Fudan’s OpenMOSS Team Open-Sourced a 5B Image Model You Prompt With a JSON Layout, So Text and Characters Land Exactly Where You Put Them MOSS-Image — Fudan’s OpenMOSS Team Open-Sourced a 5B Image Model You Prompt With a JSON Layout, So Text and Characters Land Exactly Where You Put Them MOSS-Image — Fudan’s OpenMOSS Team Open-Sourced a 5B Image Model You Prompt With a JSON Layout, So Text and Characters Land Exactly Where You Put Them

MOSS-Image — Fudan’s OpenMOSS Team Open-Sourced a 5B Image Model You Prompt With a JSON Layout, So Text and Characters Land Exactly Where You Put Them

OpenMOSS released MOSS-Image, an Apache-2.0 5B diffusion transformer prompted with a JSON layout — labels, captions and bounding boxes — with strong in-image text rendering and day-one ComfyUI support.

Issanou 21 October 2026
A Great Wall watchtower at sunset above a sea of clouds, a static 3D scene from the Marble gallery used on the OuroWorld project page
0 17 0 Comment on OuroWorld — Public Code That Makes a Frozen Gaussian Splat World Move in a Seamless, Endless Loop
OuroWorld — Public Code That Makes a Frozen Gaussian Splat World Move in a Seamless, Endless Loop OuroWorld — Public Code That Makes a Frozen Gaussian Splat World Move in a Seamless, Endless Loop OuroWorld — Public Code That Makes a Frozen Gaussian Splat World Move in a Seamless, Endless Loop OuroWorld — Public Code That Makes a Frozen Gaussian Splat World Move in a Seamless, Endless Loop

OuroWorld — Public Code That Makes a Frozen Gaussian Splat World Move in a Seamless, Endless Loop

OuroWorld turns a static 3D Gaussian Splatting scene from Marble, HY-World 2.0, Lyra 2.0 or a real capture into a looping 3D cinemagraph, with no motion mask. Code is public; the full pipeline is for non-commercial research.

Issanou 21 October 2026