"Generate a photorealistic realtime render of a human face with webGL" (Qwen3.5-122B-A10B UD-Q3_K_XL)

[-]

Bulky-Priority6824@reddit

Probably what I thinks we look like staring at these screens

[-]

Psychological-Lynx29@reddit

Shit, if we are at this point with local llm's, cant wait to see what local models can achieve in 2 more years, its going to be just like the progress Will Smith made at eating spaggetti

[-]

kant12@reddit

For Q3 that's actually better than I was expected tbh.

[-]

silenceimpaired@reddit

Flawless Victory! ... um ... Finish Him!

[-]

mr_kandy@reddit

I assume that main issue here is missing feedback loop, where resulting image is returned back for analiz to model and compared with some example

[-]

Sofakingwetoddead@reddit

Better than this amateur sketch. "anybody seen the leprechaun say yeah!"

[-]

_TheWolfOfWalmart_@reddit (OP)

Highest quant I can fit on my 4090 + 64 GB system RAM without disk swapping.

Q3 isn't as bad with the bigger models.

[-]

Force88@reddit

What is your speed with it? E.g token/s

[-]

huzbum@reddit

Meh, to be fair this is about as good as I can do too

[-]

EagerSubWoofer@reddit

you're joking, but remember that this is the least realistic that photographs will ever look

[-]

NarutoDragon732@reddit

Just one more trillion bro I swear we'll have AGI

[-]

shroddy@reddit

Can you upload the sourcecode somewhere?

[-]

_TheWolfOfWalmart_@reddit (OP)

Enjoy!

[-]

shroddy@reddit

Thx alot! I wonder why id did generate separate shader versions for browsers with and without support for WebGL 2, as both versions look the same and every modern browser supports WebGL 2.

[-]

_TheWolfOfWalmart_@reddit (OP)

Good question, I was wondering the same lol

[-]

StupidScaredSquirrel@reddit

I can't tell what's real and what isn't anymore...

[-]

_TheWolfOfWalmart_@reddit (OP)

I think AGI is here?

[-]

LetsGoBrandon4256@reddit

Quite impressive tbh considering it has to come up with all the vertices, edges and faces.

[-]

shroddy@reddit

In 3D graphics, faces are (usually) triangles in a 3D space.

You have vertices (points in 3D space), edges (connections between two vertices) and faces (usually 3 vertices / 3 edges forming a triangle, sometimes 4 vertices and edges to form a quad).

[-]

-dysangel-@reddit

has anything since the NV1 rasterised as quads? I thought everything was triangles at hardware level

[-]

_TheWolfOfWalmart_@reddit (OP)

Yeah, I mean it's terrible and funny but simultaneously impressive, all things considered.

[-]

zigzag3600@reddit

I mean, imagine hiring a boss's stupid nephew who tries his best. That, probably, what he would do – technically it's what you asked, done to the best of his abillites =)

[-]

Glazedoats@reddit

hmmm. but does the model know what topology is?? Maybe? topology for heads is brutal to learn even for humans.

[-]

BumbleSlob@reddit

Can we adopt this as the local inference community’s resident monster? Maybe we can name him Noblink

[-]

Fine_Nectarine9328@reddit

this is equivalent to will smith noodles eating video let's see how much time it will take to cover it

[-]

Marshall_Lawson@reddit

reminds me of the old smiley face billboard for Cure Insurance by the jersey turnpike

https://blogger.googleusercontent.com/img/b/R29vZ2xl/AVvXsEjzrQX9Ol-pKh8Y6bkNlaIhWi1SkyZBnynH6OWHbtLKSLiiz7lbf-CBU8Au0SNFHP6N7zOL0b27NVql5tLfL64FvjU3f5p6F026WXebd8I6dDpqGoNdFp53IdnpXnNFRELE4mrMkKIuVX_V/s1600/nj+cure+auto+insurance+reviews.jpg

[-]

LoafyLemon@reddit

Is this based on Mark Zuckerberg? I can see some resemblance...

[-]

Operation_Neither@reddit

[-]

user92554125@reddit

I suppose this is the Q8?

[-]

-dysangel-@reddit

should have asked to have it sipping some water

[-]

user92554125@reddit

LOL

[-]

Paradigmind@reddit

You should post this in r/StableDiffusion.

[-]

GrahuleDeGore@reddit

That's how we look for THEM...

[-]

lioffproxy1233@reddit

Show me one example of that being done.

[-]

Guilty_Rooster_6708@reddit

Exactly like Kevin O’Leary

[-]

dryadofelysium@reddit

FWIW this is using Gemini 3.1 Pro in Google Antigravity.

It used textures from the official three.js repo for the human:

https://raw.githubusercontent.com/mrdoob/three.js/master/examples/models/gltf/LeePerrySmith/

I can upload the project if interested. The movement in the scene is a little busted but it worked with some patience.

[-]

FullstackSensei@reddit

Gemini uses search behind the scenes a lot.

You could get the same result with probably a 9B model and a search MCP.

TBH, I don't find such tests useful anymore than the stupid how many R's in whatever-berry is. All you're doing is asking the LLM to regurgitate geometry. There's very little algorithic knowledge in such a test.

[-]

dryadofelysium@reddit

I agree but comparisons are still fun. I have a bunch of prompts saved that I want to run on the new Gemini model (3.5 I think) coming tuesday to compare.

[-]

Borkato@reddit

It used textures? That’s just cheating

[-]

_TheWolfOfWalmart_@reddit (OP)

Wow. Wait, did you supply it textures or did it come up with those itself? And how many turns to get it looking like that?

That's impressive.

I purposely tried this with no textures, nothing to start from, no planning, and one-shot. Just "hey do this right now" basically.

[-]