Matt

Blender Cycles renders of 5 of the charms in High Roller It's a dice rogue-like where you make poker-like hands with dice and beat a series of levels while you get dice upgrades and more These models are the "Charms". Things that give you effect advantages My current workflow: - @3DAIStudio to generate reference images (en-masse, using reference style images so they stay consistent) - that's GPT Image 2 - @3DAIStudio the models (@tripoai's Tripo P1 model). Ultra textures, PBR material - I modelled the table and dice tray by hand in @Blender , used the beautiful PBRs from @PoliigonHQ - The dice are a model by rogozhko on Sketchfab. I also generated the HDRi using GPT Image 2. I'm not sure if I'll use it in game but it's stunning for something out of an image model, and great for concept renders like this! I'm gonna generate a few more charms, dice and other game components Then upgrade my ThreeJS PoC and see how far we can push WebGPU / TSL to get a balance of realism, taste, fun and (no pun intended) charm

@grujicbr @TheCodeTherapy I’m fairly sure I saw a post from @thsottiaux that they were doing some stuff under the hood to reduce token spend (around the same they removed the 5 hour limit) Maybe it regressed in other areas haha

@TheCodeTherapy @MrCollison I think they broke something with SOL, 5.5 did not have this issue, just reads like to me like SOL is going crazy. I did a comparison of *asking* fable to do the same thing, and it did it with 2 lines of code. *BUT* like i said before I think 5.6 is getting rammed with usage

Played a bit with SOL 5.6 ultra with all of these limits removed. Asked it to implement a simple fix, it created a 24,000 line change diff with fully re-written a bunch of systems for no reason 🤯 What is the 5.6 equivalent of the 5.5 xhigh that doesn’t go insane?

It’s great if I give it a detailed, focused spec. I’ve already architected the higher level code at that point If I set it off running without any guidance (no report back before actually doing something + iterate) it’s just a slop cannon. All the things you mentioned + really dumb wgsl code (and im not that good at shader code myself) Fable not so much. It gets the mission and follows the pattern 90% of the time on some quite heavy stuff

I literally gave up using AI to deal with Rust + WebGPU. Too many per-frame allocations, unnecessary heap traffic, avoidable clone() usage, avoidable to_string(), format!, collect(), join(), unnecessary usage of Box, Rc, Arc, Mutex, and RwLock all over the place. Basically zero care for performance. Are you having a good experience with your tech stack using Sol?

@Art_If_Ficial @marco_fugaro Yeah, all his other demos are awesome!

@MrCollison @marco_fugaro is talented

Ugh this is so cool. I’ve been fascinated with games that have anything close to a portal-like mechanic since, well, the first portal I may have to do my own spin on this I have the perfect domain for it, but I also have several other higher priority things ongoing 😅

That’s super interesting! Do you think it’s anything to do with the code you work on evolving? I find most models, with exception of fable, do better when code is less evolved I literally had this today - Sol ripped out a bunch of features and rewrote a hunk of good rust wgpu stuff I’d hand rolled - it had basically nothing to do with the task at hand A couple days back when the codebase was less mature (this thing came together fast AF) it found and addressed a shed load of little inefficiencies and did a great job. I don’t trust it to do the same today haha. It’s kinda like AI music mastering - it doesn’t know when to stop additive work

@MrCollison It’s quite interesting but 5.6 SOL seems to just love writing so much code for the simplest things even on xhigh, but I think it’s performance is suffering from too many people using it right now. It was much much better on day 1, right now can’t trust any code from it

@SkyeSharkie That last sentence sounded AI generated - I promise it was spawned from my brain Comparisons smell like LLM speak now 😅

People are mostly great. Negative groupthink en-masse sucks. I even see it in my bubble - ignoring the learning / tooling acceleration instead focusing on how it "makes people lazy" yet, as always, the individual chooses the effort level. You'll always have copy + paste junkies & you'll always have people who wanna dig in. Pretending this is different is funny because it's almost exactly the same, it's just accelerant on who you are.

it feels weird to be one of the only 3d vibecoding poasters with actual AAA game industry experience - i was basically blackballed/shunned for liking and working with AI and now all the big companies are integrating it, the social dynamics of these industries are fucking evil, tbh

@SkyeSharkie That's pretty amazing. We need a better model that gpt image 2 / patina on the market for uv maps though because there's always so much mottling / artifacting! I wonder if a flux LORA could be trained with lots of UV (island inputs & reference image) + (output good map) pairs

okay! my whole pipeline for getting fable and sol to be able to 3d model good topology hard surface game models without using any additional paid services is WORKING... we just need to dial in their lowpoly geometry and baking techniques some more, the pipeline is images > trellis > model creates the lowpoly themselves with direct headless blender interaction > bake ... this will work with both of them because you can replace sol's gptimage2 imagegen in fable's version with z-image-turbo or another quality local image model

@JKoukourakis It looks wrong but right at the same time. All the while it's a beautiful sight regardless!

Plugged wind noise into the fog shaders to generate some movement. Interesting results. 🌫️

@mattworkman Have you turned fast mode on? It doesn't eat API based billing like Claude fast mode does!

WHY IS EVERYONE lying?! Codex is SOO slow, no one can actually be working like this!

@BRobo82 So incredibly cool. I wonder what tech LBP was built on - it had a bunch of sculpting capabilities! Probably just raw dogging mesh operations, but to think that was happening on PS3 with some pretty big levels (relatively speaking)

@MrCollison Oh yeah, the entirety of Dreams is SDF plus a point cloud on the surface. On top of that, you have flecks, which are particles and are also made of point clouds. You can illuminate all sculptures and flecks this way with zero performance hit.

Something similar exists in Dreams, and it costs practically zero performance. You can practically paint with light at no performance cost, except for memory.

Have been optimizing to run on Apple TBDR and mobile. No hardware raytracing but custom compressed binary BVH with TinyBVH for coherent rays. Timings are per-frame to compute a single bounce from scratch, no temporal accumulation. Apple M2 baseline 1x / 2x (supersampled)

@CodyBennett @majidmanzarpour @threejs @garrettkjohnson That's a golden test case. Thank you! I can finally test the local irradiance volume implementation this weekend. The work week can't go fast enough!

@MrCollison @majidmanzarpour @threejs @garrettkjohnson Amazon Lumberyard Bistro is a good one I think.

idTech 8 style Global Illumination (FAST AS HELL) in @threejs (via WebGPU) I've worked on this for 5 weeks (on and off) and it's still not there, but all the core tenets work. Next more denoising + better temporal accumulation, and some larger scale examples. (Running at 4K)

@CodyBennett @majidmanzarpour @threejs @garrettkjohnson My intuition exactly. Also, I wondered what it meant when I saw "external example"! I will find a way to publish it, portability is gonna be the biggest challenge so just example scenes to start with. Do you know of any big outdoor+indoor scenes to stress test with? Even paid?

@MrCollison @majidmanzarpour @threejs I don't think this would be tenable for Three.js to maintain as an example but rather as something external. If you publish it somehow, Three.js has incorporated external examples lately such as @garrettkjohnson's work. Or better, perhaps it could be showcased on the website.

"Dynamic Diffuse Global Illumination with Ray-Traced Irradiance Fields" jcgt.org/published/0008… "Scaling Probe-Based Real-Time Dynamic Global Illumination for Production" jcgt.org/published/0008… - not cited in Tiago's talk BUT its an extension and I reckon it was used "Real-Time Global Illumination using Precomputed Light Field Probes" research.nvidia.com/sites/default/… - again not in his talk but 5.2 moment visibility in a DDGI like way - no precomputation There's a lot more but I need to have lunch before going back to work 🫡

I can't remember the exact point light leaking didn't become an issue; there are a ton of small members iirc when I was reading / implementing parts of Chebyshev visibility + trilinear interp across neighbouring probes that made a big difference Also in one of the scaling DDGI papers - using backface hits / storing zero irradiance to disable LIKELY trapped probes, but there are nuances to the techniques I didn't manage to implement. E.g. I imagine idTech 8's final one implemented probe relocation + a custom DDGI version of "distance-statistics" from Light Field Probes paper There is so much more context to share - I should probably just ask GPT to dump out a table for you but I don't think it would know the paper references Let me get some links Oh and on the surfels / radiance caching paper defo a Tiago answer on that. I know Surfels is cited in prior research inspiration but not in the disclosed implementation. Radiance caching paper looked too memory heavy when I looked at it (and distinctly different pipeline in the methods - portable bits but I am no way near experienced enough to give it a go)

@ImLunaHey @thsottiaux We get a banked reset today 🤩

im really surprised with just how much ive gotten done with a $20/m sub to codex. in saying that... @thsottiaux can we pretty please have a reset soon?

@gfodor Bright stuff in my kitchen is literally what inspired me to look into GI in the first place 😂

@MrCollison anyone who has owned a bright pink fridge recognizes when GI is left out

This scene shows both GI & raytraced reflections clearly. It's also great for debugging noise because it's fairly flat and uses only one directional light. This scene would be basically unlit without GI, yet look at the indirect light below the fridge & the counter reflections!

@SilasCodes3D @threejs It does look super cool right 😎

@MrCollison @threejs No idea what any of that means bit it looks super cool 😮🫪🚀