Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

Bonus: the same model does metric depth completion. Give it sparse lidar points, it fits a small LoRA at test time and fills in the rest, no retraining. This is similar to the Marigold-DC style guidance loop, but simpler. (Video from Marigold-DC)

Anton Obukhov

3,063 subscribers

12,256 views • 1 month ago •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

Here's the new American model. Devil is still in the details between the EURO and the GFS. The GFS has the heaviest snows DC & North...while the European has it just south of DC. Which are you rooting for? (Totals from this run below this)

Here's the new American model. Devil is still in the details between the EURO and the GFS. The GFS has the heaviest snows DC & North...while the European has it just south of DC. Which are you rooting for? (Totals from this run below this)

Mike Thomas

605,324 views • 1 year ago

i mentioned this in the comments but ktisis does have invisible skin feature built in :3 it does however apply to the entire model but nothing it stopping you from creating clones via brio and doing the same But if you need precise control I would recommend still using penumbra

i mentioned this in the comments but ktisis does have invisible skin feature built in :3 it does however apply to the entire model but nothing it stopping you from creating clones via brio and doing the same But if you need precise control I would recommend still using penumbra

peachy!! (modding prog)

65,867 views • 1 month ago

This is native Qwen Image 2.1, no LoRA - move and resize objects seamlessly in an image🔥 I was planning to finetune on this task to showcase the new 🧨 diffusers LoRA support for Qwen 2.1 (go try it!), but while testing the base model I realized it's actually so good at it out of the box (pun intended haha) Try it ▶️ Train 🧨

This is native Qwen Image 2.1, no LoRA - move and resize objects seamlessly in an image🔥 I was planning to finetune on this task to showcase the new 🧨 diffusers LoRA support for Qwen 2.1 (go try it!), but while testing the base model I realized it's actually so good at it out of the box (pun intended haha) Try it ▶️ Train 🧨

Linoy Tsaban

52,617 views • 13 days ago

I'll always root for a team that open-sources its best work, and Robbyant just did it properly. Robbyant, Ant Group's embodied-AI company, released LingBot-Vision, a vision foundation model for robots, and the part I love is the data. They trained it on 161M images, filtered down from 2B raw ones and mostly pulled straight from the open web, with no human labels, no edge detectors, no depth sensors anywhere in the loop. It learns the exact edges of objects from raw pixels. That's roughly a tenth of the data DINOv3 saw, and under a third of the training. And it shows in the results. On depth, working out how far away things are, the 1B model edges out a 7B on NYU-Depth. It also powers LingBot-Depth 2.0, which reads the surfaces cameras usually choke on, glass and mirrors, and halves indoor depth error. LingBot-Vision is fully open. Weights from the 1.1B flagship down to a tiny 21M version, code, and the paper. This is the timeline I want more of. Robbyant

I'll always root for a team that open-sources its best work, and Robbyant just did it properly. Robbyant, Ant Group's embodied-AI company, released LingBot-Vision, a vision foundation model for robots, and the part I love is the data. They trained it on 161M images, filtered down from 2B raw ones and mostly pulled straight from the open web, with no human labels, no edge detectors, no depth sensors anywhere in the loop. It learns the exact edges of objects from raw pixels. That's roughly a tenth of the data DINOv3 saw, and under a third of the training. And it shows in the results. On depth, working out how far away things are, the 1B model edges out a 7B on NYU-Depth. It also powers LingBot-Depth 2.0, which reads the surfaces cameras usually choke on, glass and mirrors, and halves indoor depth error. LingBot-Vision is fully open. Weights from the 1.1B flagship down to a tiny 21M version, code, and the paper. This is the timeline I want more of. Robbyant

Chubby♨️

48,249 views • 3 months ago

I made a little tool to help me plan DC Metro transfers and pick the right car to ride in. It uses real-time data from WMATA I'm going to release it to the public for a bit, feel free to give it a spin, it works great on mobile too

I made a little tool to help me plan DC Metro transfers and pick the right car to ride in. It uses real-time data from WMATA I'm going to release it to the public for a bit, feel free to give it a spin, it works great on mobile too

hazel☀️

14,934 views • 8 months ago

When I tell you this Usher and Chris concert is whack at the National’s stadium in DC, it is. Spent a few grand for this view.

When I tell you this Usher and Chris concert is whack at the National’s stadium in DC, it is. Spent a few grand for this view.

JANEE RAE💋

124,281 views • 3 months ago

this tiny solar tracker tilts its panel toward a moving light source two LDRs sense the imbalance. a small geared motor does the rest it’s a rough DIY build, but the final flashlight test is a lovely proof of the idea want a guide to build the same prototype?

this tiny solar tracker tilts its panel toward a moving light source two LDRs sense the imbalance. a small geared motor does the rest it’s a rough DIY build, but the final flashlight test is a lovely proof of the idea want a guide to build the same prototype?

ard

29,522 views • 1 month ago

A couple of months later, in a very similar situation, but this time, Balde is calm enough to not only see the pass but, more importantly, execute it properly and not rush it. The boy is learning.

A couple of months later, in a very similar situation, but this time, Balde is calm enough to not only see the pass but, more importantly, execute it properly and not rush it. The boy is learning.

sm

759,377 views • 3 years ago

Don't train the model, evolve the harness. I read a brilliant blog post from Hugging Face where they took a frozen open model scoring 0% on a hard legal agent benchmark, left its weights alone, and let an automated loop rewrite only the code around it. That code layer is the harness, the runtime wrapper that feeds the model context, runs its tool calls, and decides when a run ends. By the time the loop finished, the system had essentially matched Sonnet 4.6 on the benchmark's headline metric, at roughly 7x lower cost per task. Zero weights changed. The gain existed because of where the model was failing. The judge only grades files saved in the right place under the exact requested filename, and the model kept doing the legal analysis correctly, then saving it under the wrong name, dropping it in a scratch folder, or never writing it at all. So the 0% was never measuring legal reasoning. It was measuring the harness. Hand-tuning that layer is slow and model-specific, so they automated it. A Claude proposer adds exactly one mechanism per iteration, and an outer loop keeps it only if it clearly beats the current best, so accepted mechanisms compound. What the loop discovered says a lot about where agents actually fail. → The biggest single gain was file handling, not intelligence. An automatic step that lands the deliverable exactly where the judge expects it beat every prompt change, with zero extra model tokens. → Code fixes transferred across models, prompt playbooks did not. The same harness lifted a smaller model from the same family by 14 points, but the tuned prompts hurt a different model family on tasks it could already finish. → The harness mattered more than anything else. Same model, same judge, same tasks, and five different harnesses scored anywhere between 3.5% and 80.1%. The gains do eventually flatten, and the remaining misses look like real capability gaps. At some point the wrapper runs out of tricks and the model has to carry the work. But the lesson holds. A benchmark score measures the model and its harness together, and until the harness is fixed, it's impossible to know which one failed. I highly recommend reading this: I also wrote a deep dive on agent harness engineering a while back, covering the orchestration loop, tools, memory, context management, and everything that turns a stateless LLM into a capable agent. The article is quoted below.

Don't train the model, evolve the harness. I read a brilliant blog post from Hugging Face where they took a frozen open model scoring 0% on a hard legal agent benchmark, left its weights alone, and let an automated loop rewrite only the code around it. That code layer is the harness, the runtime wrapper that feeds the model context, runs its tool calls, and decides when a run ends. By the time the loop finished, the system had essentially matched Sonnet 4.6 on the benchmark's headline metric, at roughly 7x lower cost per task. Zero weights changed. The gain existed because of where the model was failing. The judge only grades files saved in the right place under the exact requested filename, and the model kept doing the legal analysis correctly, then saving it under the wrong name, dropping it in a scratch folder, or never writing it at all. So the 0% was never measuring legal reasoning. It was measuring the harness. Hand-tuning that layer is slow and model-specific, so they automated it. A Claude proposer adds exactly one mechanism per iteration, and an outer loop keeps it only if it clearly beats the current best, so accepted mechanisms compound. What the loop discovered says a lot about where agents actually fail. → The biggest single gain was file handling, not intelligence. An automatic step that lands the deliverable exactly where the judge expects it beat every prompt change, with zero extra model tokens. → Code fixes transferred across models, prompt playbooks did not. The same harness lifted a smaller model from the same family by 14 points, but the tuned prompts hurt a different model family on tasks it could already finish. → The harness mattered more than anything else. Same model, same judge, same tasks, and five different harnesses scored anywhere between 3.5% and 80.1%. The gains do eventually flatten, and the remaining misses look like real capability gaps. At some point the wrapper runs out of tricks and the model has to carry the work. But the lesson holds. A benchmark score measures the model and its harness together, and until the harness is fixed, it's impossible to know which one failed. I highly recommend reading this: I also wrote a deep dive on agent harness engineering a while back, covering the orchestration loop, tools, memory, context management, and everything that turns a stateless LLM into a capable agent. The article is quoted below.

Akshay 🚀

245,379 views • 3 months ago

China just dropped an open source model that turns any phone into a 3D scanner you point it at a room and walk through it the whole space builds itself into a 3D point cloud in real time as the camera moves, no LiDAR and no depth sensor anywhere it held a full 13 minute walkthrough at 25,000 frames without the map drifting or collapsing > ~20 fps on a single GPU > works indoors and outdoors from plain footage > the scan normally needs a $50,000 rig and a technician > Apache 2.0 and runs fully offline on your machine repo in the reply

China just dropped an open source model that turns any phone into a 3D scanner you point it at a room and walk through it the whole space builds itself into a 3D point cloud in real time as the camera moves, no LiDAR and no depth sensor anywhere it held a full 13 minute walkthrough at 25,000 frames without the map drifting or collapsing > ~20 fps on a single GPU > works indoors and outdoors from plain footage > the scan normally needs a $50,000 rig and a technician > Apache 2.0 and runs fully offline on your machine repo in the reply

savip.

105,419 views • 2 months ago

This video from the Epstein files appears to be taken on an island (Perhaps Epstein island). In it, the DOJ attempted to redact a man's face but failed. Does anyone know who this man is at the end of the video?

This video from the Epstein files appears to be taken on an island (Perhaps Epstein island). In it, the DOJ attempted to redact a man's face but failed. Does anyone know who this man is at the end of the video?

Ed Krassenstein

10,228,685 views • 8 months ago

Happy wheels is somehow SLOP and PEAK at the same time. This level took me WAY TO LONG to beat but it was FUN and CHARMING the whole time.

Happy wheels is somehow SLOP and PEAK at the same time. This level took me WAY TO LONG to beat but it was FUN and CHARMING the whole time.

Windpress

42,557 views • 15 days ago

Okay so someone suggested trying WAN 2.1, a video model that supports LoRas So I trained a LoRa of myself and it works really really well, super high resemblance, obviously you can't use a Flux LoRa but similar effect, it just works great A few annoying things: - WAN 2.1 FPS is like 16 fps, so jittery, so you have to run it through an interpolator to add another frame to get close to 30fps - Limited to about 5 seconds clips (but that seems to be a bug, see "Bug Fix I2V Frame Count Hardcoded to 81 (!!) and Optional Batch Mode") Here it is with WAN LoRa + my ElevenLabs voice + 2x interpolation + boomerang loop to make it from 5 into a 10 second video Very big potential this is the working solution if I can make this pipeline work 🤯🤯🤯🤯🤯🤯

Okay so someone suggested trying WAN 2.1, a video model that supports LoRas So I trained a LoRa of myself and it works really really well, super high resemblance, obviously you can't use a Flux LoRa but similar effect, it just works great A few annoying things: - WAN 2.1 FPS is like 16 fps, so jittery, so you have to run it through an interpolator to add another frame to get close to 30fps - Limited to about 5 seconds clips (but that seems to be a bug, see "Bug Fix I2V Frame Count Hardcoded to 81 (!!) and Optional Batch Mode") Here it is with WAN LoRa + my ElevenLabs voice + 2x interpolation + boomerang loop to make it from 5 into a 10 second video Very big potential this is the working solution if I can make this pipeline work 🤯🤯🤯🤯🤯🤯

@levelsio

126,971 views • 1 year ago

For the facial animations in my game, I use a technique that yields results similar to L.A. Noire. To keep it simple: I first generate a facial animation video using AI like LTX 2. Then, I create a video depth map from that animation and project it onto a face mask. Using vertex displacement, the depth map dynamically deforms the face as the character speaks. The result looks good in standalone VR. I actually created this workflow long time ago:

For the facial animations in my game, I use a technique that yields results similar to L.A. Noire. To keep it simple: I first generate a facial animation video using AI like LTX 2. Then, I create a video depth map from that animation and project it onto a face mask. Using vertex displacement, the depth map dynamically deforms the face as the character speaks. The result looks good in standalone VR. I actually created this workflow long time ago:

Alex

48,582 views • 7 months ago

Issue 4 of Interfaces just released! This issue is about building an animated sign-in dialog. It covers everything that goes into building it, from the animations to the small but important details. It's one of the most in-depth issues yet, full of code snippets and demos.

Issue 4 of Interfaces just released! This issue is about building an animated sign-in dialog. It covers everything that goes into building it, from the animations to the small but important details. It's one of the most in-depth issues yet, full of code snippets and demos.

Jakub Krehel

37,748 views • 4 months ago

Anyone saying that fixing DC is a waste of money does not love America or care about our history. I’m excited to see it again in all its glory! I haven’t been there in a couple of years but the last time I saw it sparkle was when I was a kid. This is so exciting!

Anyone saying that fixing DC is a waste of money does not love America or care about our history. I’m excited to see it again in all its glory! I haven’t been there in a couple of years but the last time I saw it sparkle was when I was a kid. This is so exciting!

Playteaux

54,238 views • 4 months ago

Most AI video tools are stuck in a loop, you generate, change the prompt, and wait again. Vidu S2 is pushing things forward by making generation interactive. You can edit the video while it is being created. Change the style, swap the background, or switch outfits in real time. If you drop in a new image, the video adjusts instantly without awkward cuts. Under the hood, it combines real time 720P output with an AI agent that actively watches the video to plan what happens next. It allows you to change reference images on the fly, powered by a model optimized specifically for live streaming. This opens up massive possibilities. Imagine viewers controlling visual changes during a livestream, people transforming into game characters, or shoppers trying on different looks live. The entire S Series is available for a free trial right now so you can test it yourself. Use my code PANDAS2 with the link below to grab 300 bonus credits to explore the other features. Vidu Stream: 300 Bonus Credits: Vidu AI #ViduAI #ViduS #ViduS2

Most AI video tools are stuck in a loop, you generate, change the prompt, and wait again. Vidu S2 is pushing things forward by making generation interactive. You can edit the video while it is being created. Change the style, swap the background, or switch outfits in real time. If you drop in a new image, the video adjusts instantly without awkward cuts. Under the hood, it combines real time 720P output with an AI agent that actively watches the video to plan what happens next. It allows you to change reference images on the fly, powered by a model optimized specifically for live streaming. This opens up massive possibilities. Imagine viewers controlling visual changes during a livestream, people transforming into game characters, or shoppers trying on different looks live. The entire S Series is available for a free trial right now so you can test it yourself. Use my code PANDAS2 with the link below to grab 300 bonus credits to explore the other features. Vidu Stream: 300 Bonus Credits: Vidu AI #ViduAI #ViduS #ViduS2

AI Panda

98,066 views • 25 days ago

Traction force microscopy is challenging. A less quantitative but simpler approach is to analyze the alignment and remodeling of matrix fibers around contractile cells. Does it actually work, and how? Take a closer look at

Traction force microscopy is challenging. A less quantitative but simpler approach is to analyze the alignment and remodeling of matrix fibers around contractile cells. Does it actually work, and how? Take a closer look at

Ben Fabry

22,574 views • 3 years ago

I'm 99% sure Slick Boots is bugged. I felt NO difference at all in a Ranked game and went to test it in the training cave. Colt crosses the finishing line basically at the same time with both SP Frank™ KEIENBURG 🏳️‍🌈 Dani Brawl Stars

I'm 99% sure Slick Boots is bugged. I felt NO difference at all in a Ranked game and went to test it in the training cave. Colt crosses the finishing line basically at the same time with both SP Frank™ KEIENBURG 🏳️‍🌈 Dani Brawl Stars

Kay BS

77,007 views • 1 year ago