Dragos Silion

Cofounder at @ArtlinkAR
ColorfulCoding
Doing all kinds of AR/VR projects.

[Dragos Silion @silidragos]

Soo, finally done with the AI Face Animator. All the code is public, if anyone is curious to check how this works.

Client app: https://t.co/jhnfD4cIOw

[Visual Atelier 8 @visualatelier8]

Projection mapping visual by dfifty

[Dragos Silion @silidragos]

One-shot learning works best when examples are rare:

– Siamese networks - used for FaceID, signature matching etc.
– Few-shot image classifiers (like Matching Networks or Prototypical Networks)

[Dragos Silion @silidragos]

Heard of zero-shot or one-shot learning but never quite nailed the difference?

🔹 Zero-shot = the model handles a task it’s never explicitly seen
🔹 One-shot = it sees one example and learns from it

Here’s a few very popular examples 👇

[Dragos Silion @silidragos]

Zero-shot in action:

– CLIP → finds images from text prompts without training on the categories
– Segment Anything (SAM) → segments objects from just a point or box
– GPT/LLMs → solve logic puzzles without training examples

The key: no task-specific fine-tuning.

[Dragos Silion @silidragos]

What’s next?

Gaussian Splatting is already being adapted for shading, ray tracing, and animation.

It’s fast and practical. But this field moves quickly.
I wonder if splats the new standard, or just a stepping stone to something better?

Hard to say. But the momentum is real 🧑🏻‍🚀

[Dragos Silion @silidragos]

NeRFs were a major breakthrough in 3D reconstruction.

But volumetric rendering is slow. Tracing rays through a field isn’t ideal for real-time use.

3D Gaussian Splatting changed that by switching to rasterization. No rays, just points, rendered fast.

[Dragos Silion @silidragos]

Why Gaussian Splatting is a big deal:

1. Real-time rendering (it runs directly on GPUs)
2. No neural net at inference - the scene is baked into point data
3. Easy editing & compositing. Splats behave like graphics primitives.

It’s efficient, simple, and surprisingly sharp.

[Dragos Silion @silidragos]

Must-read papers:

Latent Diffusion Models : https://t.co/bCwXFwnlkT
Latent Priors (CVPR 2023): https://t.co/CAFdHXSNjY
CVPR 2025 latent video gen: https://t.co/vSa8UbDT9T

Worth every scroll.

[Dragos Silion @silidragos]

What is Latent Diffusion?

Instead of generating images pixel by pixel, the model works in a compressed “latent” space — like sketching the idea first, then decoding it.

But here are 5 things about it that blew my mind 👇
#AI #LatentDiffusion

[Dragos Silion @silidragos]

4. Noise becomes a knob in latent space, it bends meaning, not just visuals

5. Tiny LoRAs work great just because latents are dense with structure

[Dragos Silion @silidragos]

🔧 Try it yourself:

→ GitHub: https://t.co/3cmPKDQKtI
→ Pretrained weights ✅
→ Simple command line run
Precision upscaling, creator-ready.

Follow for more tools from #CVPR2025.

[Dragos Silion @silidragos]

📈 Want control over how AI upscales your images?

PiSA‑SR was dropped during #CVPR2025 a diffusion model that lets you balance pixel-level sharpness & high-level structure.

One-step. Fast. With code.👇
#SuperResolution #AI

[Dragos Silion @silidragos]

Most upscalers split pixel & semantic detail into separate models or stages.

PiSA‑SR skips the complexity. It uses Classifier Score Distillation (CSD) to guide the semantic LoRA—no extra model, less memory, cleaner output.

Smart signal > stitched hacks.

[Dragos Silion @silidragos]

The future of generative video isn’t just better quality.

It’s speed.
It’s scalability.
And it’s models that finally understand:
“A cat jumps onto a table and looks around.”

Link: https://t.co/i1COONmE1L
Follow for more breakthroughs from #CVPR2025.

[Dragos Silion @silidragos]

(thread 🧵)
🎬 Ever typed a sentence and got a video back in seconds?

That's text-to-video generation.

But until now, it’s been slow, clunky, and limited to a few seconds.

Here’s why that’s changing – and what “linear” models are doing to speed things up.
#CVPR2025 #AI

[Dragos Silion @silidragos]

Why this matters:

🎥 Longer, smoother videos
💻 Lower hardware costs
🧪 Real-time applications (think: AI filmmaking, XR avatars)

Linear text-to-video is a building block. Not perfect, but the direction is clear.