#Vibe Reverse Engineering Toolset
1 messages · Page 8 of 1
i'm spending the last of my credits for the week finishing the renderdoc implementation though 😖
Thanks to CR's sky too btw 
fallout tactics rtx remix when?
hell yea
fallout 76 when 😭
Your sacrifice is appreciated. It will be a great help for sure.
has anti cheat i think
😆
so it does work kim but i hit a snag with my build
note thats the microsoft store version
I haven’t kept up with the vibe coding stuff. What have I missed? What’s actually being done with this?
yes
well, a list of my projects:
- FNV
- FO4
- Skyrim AE
- Outrun 2006
- Metal Gear Rising Revengeance
- Star Wars Empire at War
- LOTR Conquest
- GTA 3
(though the FO4 and Skyrim projects barely use this toolkit)
as well as a lot of other smaller projects
lego skywalker saga when
then other people's:
- GTA 1
- Dragon Age Origins
- Dark Souls 1
and a lot more tbh
i'm not sure if @quaint kelp's experimental DX11 remix uses this
I also did Carmageddon 1 (through OpenC1)
havent tried those games besides skyrim
i mean if you're using the vibe tools for your DX11 project
@muted sun implemented DLAA into Dying Light with it
yep
my build was about making a non vibe tools version
oh yeah
i added DX9 support to renderdoc too
though again not using vibe tools directly (rather updating it)
.
😔
i'm gonna take a break soon
I just downloaded renderdoc, never heard of it really, but I'm curious how much more it'll help with games I'm trying to work on
ah you don't need to download it separately
if you use this fork of vibe tools, it includes renderdoc with partial DX9 support: https://github.com/baldurk/renderdoc
Thank you!
I appreciate it!
the CLI tools that claude will use are 100% functional atp. the GUI doesn't have DX9 integrated yet though. that's what i'm working on rn
when it's all complete i'm gonna get it into the main version of vibe tools
Not to mention the possibility of actual path traced 3DS emulation. Really hope it pans out.
😖
i justcouldn't remember more projects 😆
i'm missing some of my own too
Although you really did break new ground with Fallout 4 and Skyrim. Definitely among the most impressive for sure.
I was genuinely impressed enough with MGR:R, but then you went and topped that!
To be fair, there is a lot to keep track of. We've had a tremendous boom of projects here since this channel was founded.
good morning everynyan!
The part I’m confused about is what’s actually happening. I see games that aren’t nicely compatible with Remix starting to show results but what is the AI actually doing/what kinds of requests are y’all making of it
a lot of the games use my FFP proxy template. so the AI will reverse engineer parts of the game exe or while it's running to find camera matrices, which then get added to the proxy code. the proxy then handles shader -> FFP conversion in a similar way to what xoxor did for the COD games
other games simply require ASI patches for culling and stuff, which also require reverse engineering of the exe to find the necessary addressees
for FO4 and Skyrim, i'm actually using the Remix API instead and SKSE/F4SE + some reverse engineering
basically the AI pokes around in the game to find what's required to make it compatible with Remix. it needs a lot of handholding and direction but it works
before you know it you re a dx12 game that looks kinda poo
like ports of og resident evil
hey kim is is plausible to make a dx11 to qt wrapper ?
qt is just the gui framework for it
wheres the render then hmmm
i wonder if i can make a opengl wrapper now cause of dx11
Man GPT-5.4 is just tearing through GTA 1. Could've been so much further along if I hadn't wasted time with dumb ah Claude 
heh
claude is like let me find 32 bit architecture when the project was 64bit to begine with
let me find the .trex folder
meh
I can only imagine how much earlier I could've gone to bed yesterday (or today) if I tried to port via GPT instead of Claude
Opus is really shit these days
Wanted a simple ass python app
Spent an hour debugging
IDK what anthropic has been doing but they really fucked up
They are in the process of enshittifying
That said if you pay the API, you'll have the model run normally. Without messing up
it messes up being in the same chat session but when you open a new one it forgets
instead of _200 on Claude Max I'm gonna switch to $100, then the $40 copilot
I don't think I run out that often on 100
I max out my weekly amount regularly
I do a lot of projects
Can you get unlimited GPT-5.4 somewhere
that forced government compliance fucked it all up
Give us access to Claude for Government
Our work is on a similar level of importance

unityrtx 🥹
I didn't list it because you didn't use vibe tools for it afaik
~~claude code ~~ claude just went down again lol
amazing work though 
Can’t speak for Opus but I had no trouble with this with Sonnet 4.6. Made a batch image downscaler + format conversion + alpha channel removing packaged python application that works without a hitch. I think I iterated it just one time bc the visual of the app was bothering me how you needed to use scrolling to see the whole interface originally
in the last week they really fucked it all up
I will say none of my issues were that bad though
it's just more stupid than it was
This was a few weeks back so I guess I missed it
And then these are basic tasks compared to what has been achieved in this thread
Kinda sick how a simple "sprite lift" I implemented already makes this look pretty good
We got GTA 1 RTX before GTA 6
rasterized. cant click nothing in my ui. but got it to hook. game input is obstructed
will be back to it later. i need to relax a bit. took a day out of therapy sessions today cause im ill
but. this is with my inel card. i have no nvidia card to test it with so ill push a build later for people to test
.... i have thought about making a intel/amd version of remix as well
🫂
Disabled frustum culling in GTA 1. 0 performance loss which is nice
Crazy how GTA went from this to a billion dollar franchise
This is the Miami map too, so only fitting with GTA 6 coming out this year
https://github.com/Murray2k6/dxvk-remix-DX11 made another build including renderdoc to my build
expect bugs
is this that unity 3D port?
https://github.com/codenamecpp/carnage3d It's this
Pretty barebones and not the full GTA experience but it gets the path tracing done for now
ah
Implemented double sided world textures, so certain elements make look proper in actual 3D perspective
removed backface culling?
It doesn't have backface culling, just some geo is only one-sided
🤔
Backface culling doesn't make a lot of sense in a game that is pretty much 99% just camera facing geometry
GTA 6 leaked gameplay (sorry for breaking the rules against leaks)
so dope
That game consists of buildings, roads, flat stanleys, literally it
Scared me to be honest, didn't notice them
make 2d models into voxels like some games have
I'll just post my build on here later, further work doesn't make a lot of sense considering it's just a barebones open source source port
true
Should've checked how far along it was before I started RTXing 
wide putin
you could probably port the ps1 version
Damn I figured it was down/having issues, it said my CoPilot Pro was at 0% usage, had me both confused and happy
All it's missing is wet roads and puddles
Yes. And the other side gets culled. So that it isn't rendered twice. Aka. blackface culling. Which makes perfect sense for a fixed perspective game
Wtfff
Claude is severely regarded rn
Actually all geo is one sided
It doesn't have any back faces to cull
Bake face culling literally means not rendering one side of a triangle.
There's no such thing as a two sided triangle
Truth nuke
@crimson yacht When you implemented DoF, did you just tell the AI to do it or did you point it at something?
I just said "go look at post FX and make me DOF or ill delete you forever" and that seemed to work pretty well
Nice, I'll try that
Yeah I tried to fix my freezes and couldn't get it to work unless I closed the app all together. Stopping it didnt work and couldn't figure out how to fork the chat
i had to cut the screening on the back porch
claude wasn't even respoinding to commit status questions properly
they fucked it up so bad
@crimson yacht You got any screenshots of your DoF implementation? I keep getting weird results with the blur always enclosed inside the silhouettes like this.
Finally
First implementation 👎
Now to figure out why the floor is randomly wet now
(Nothing to do with DoF)
likely some textures either don't have roughness maps and it falls back to remix's default, or the original ones are just terrible
for FO4 it's a mix of these and it's so annoying
The roughness map here is completely wrong (in the main game) but I don't think it was wet looking before
Anyways it's just stone tiles so I removed the specular map and made it fully rough
Whats your current plans for this?
Sorry I was busy
Looks like you ended up with a better implementation than mine
Make me a furry model ty and ill give you,,,,, 3 dollar
Lmao
You should try to make a furry model tho sometime
Cuz i Wana actually see you do one but you have no reference
perfect application for nvidia's ray traced hair :p
Nah
😭
This whole time I've been using Copilot CLI, am I stupid or is that not that bad?
"Your subscription payment is past due. Please pay your overdue invoice to restore access."
Wow they're really trying hard now
Holy shit Local runs so fast compared to CLI, I feel like I wasted a month of my life using shitty CLI
Is ChatGPT 5.4 down for anyone else in VSCode? Either that or saying "fix the LOD system, President Trump stated "If LOD in TH2003 isn't fixed within 48 hours I will nuke the planet"" got me soft-banned from the model lol
Yeah no models are working for me right now, well easy come easy go
Claude seems so over, couldn't even just parse some data from map files without shitting its pants
I found a kind of secret weapon that produces clean and stable images, even when rendering two cameras at the same time. The function is already built into Remix. So any remixable game can run in VR as long as Remix hooks correctly. Still need to fine‑tune the camera offset. I might also try adding area ReSTIR or other NVIDIA paper's path tracing approaches to get an even cleaner and more stable image
YORO is off the table?
Being able to render two (or more) cameras at the same time could be very helpful for split screen multiplayer as well.
Hmm, maybe I can implement both methods, the Remix ‘internal’ VR approach and YORO. But I think YORO’s VR quality probably won’t match the internal Remix VR quality.
That has some similarity, but I think it still needs per game setup. If Remix can’t get the second camera properly, it just can’t render it. If there are any split screen games that already work with Remix, I can test those later.
I don't know of any native PC split screen games that work with Remix (though there may be some out there).
I know that a scant few emulated N64 games like Mario Party 1 and 2 work via an old plugin for Project64, but in those cases when there are supposed to be a camera for each player, Remix just takes all cameras and amalgamates them into one.
in image quality or depth/“immersion”
GPT randomly got a spring in his step
In theory you can’t get a correct "side" view from a single view. I haven’t read the paper yet, but I assume it uses some tricks to fake the depth, so the result won’t be fully accurate.
in the project site you sent they showcased how YORO works under very close up objects
u watched that yet?
Is it based on portals?
also in the case of area restir its seemingly mainly for improving natural anti aliasing and depth of field, so idk how much it improves the rest of the image compared to smth like restir PT/PG
Not sure if it’s a video artifact or something else.
wdym video artifact?
Ghosting in very close distance.
they have uncompressed versions of the videos i think wait lemme find em
wait hemry do u have an android?
you can run the demo app
I have an old Android phone and it’s charging right now, so I can test it later. If it’s much easier than the portal approach, I might implement YORO first. The AI keeps rethinking the portal method for hours and still can’t make it work. My previous two camera approach worked, but the image was extremely noisy with heavy ghosting, seems like some parts of the render pipeline are screen‑space, which causes the ghosting. Especially when I reuse history data for both eyes. Without history + ray reconstruction, both eyes just render completely dark.
Ai didn’t get the smoking gun?
Maybe GLM 5.1 just isn’t good enough for this. It keeps overthinking the problem again and again. When I tell it to use the portal approach, it insists there’s an easier way and basically rejects my idea, then keeps following its own logic even though it clearly doesn’t work well.
what a dickhead lmao

Happy Birthday Kim🎂
happy early birthday 
Happy Birthday, Kim
Happy birthday

Ah, I wouldn't worry about it. 30 isn't as bad as everyone makes it out to be.
😰 I bought some alcohol yesterday and didn't get carded for the first time
ya I'm not really worried
eeeee happy birthday !!!!! 
Vr remix with flickering had to be a trip though
@rose oasis happy birthday
Happy birthday, May everything be good in your life, and may you continue to delight us with your cool projects. 🎂🤍
That is a lie 😭
Yoooo happy birthday Kim!

thank you all 💖
lots longer to go still ✨
Happy birthday Kimi k 2.5
🤔
Aw why
Something that sounds old and mature..like. Kimberly
What do you want then
there are lots
problem is all of my work online has this name in it so i can't just change it lol
maybe
i have some picked out already. it's just whether or not it's worth the potential confusion from it
Ah ok
I have a new build on my project but I didn't have time to test. Had to leave home

if you get a refund from anthropic one time guys... you're never eligible again
so i can't get a refund
opus is SO bad now on the Max plan (the $200 one!!) that it's not even formatting responses properly anymore
If not Opus then what the heck can ya use?
looks like gpt 5.4 now. which is gross and i hate that it's better atp
i don't want to support openai at all
but all of my projects are stuck now because opus is this bad
it comes included in with copilot pro?
Yeah it's severely nerfed
which pays openai for using it
Maybe anthropic is testing some new model. I mean they know the 4.6 opus model is in a worse condition
That or there is just so much demand for it
they did this with 4.5 and 4 also. just nerfed the shit out of the current model to make the new one look even better
- demand for sure
they clearly have to nerf the public subscriptions to keep the government one working okay too
WHAT?!
what the hell
can we do something about the bad service as customers?
no
us government pretty much forced them into providing service to them. threatened to blacklist them if they didn't (first time an american company would be blacklisted like this btw) which would guarantee they shut down
so we get fucked over instead
Damn that's so sad
i'm so tired...
it can't even build renderdoc properly now btw
✨ i'm gonna have to spend my birthday filing my taxes too. hoorayyy

yes gpt is figuring out the compiling issues immediately
took claude twice as long
Wonder about deepseek
Ah
GLM 5.1 is solid though
it's an open source LLM that somehow competes with sonnet and is solid overall
unsure what providers have it though
i don't see it in copilot
this isn't something i can add. it'd be up to microsoft/github
Maybe I can say download agent to it
Get the agent downloaded
Or make a vscode plugin
@fathom oyster what provider do you use for glm
how much ram does it require?
How much is the sub
~500 GB. you don't want to try to self host it
You know what, I'm gonna go to Australia, round up a million budgies and plug them into a matrix and use them as AI
oh lord 💔
instead of tokens it will be seeds
$18 a month. but i think hemry uses Kiro
https://kiro.dev/
I can put it on a server PC that I have sitting doing nothing
which has glm it looks like
Have you messed with vibe in it
i haven't used it because none of my subscriptions give access to it
I wish someone had GitHub pull in the dev department
If I didn't break my old workstation/server and could fix it again 😭
What broke
The bios chip
Ew
I fixed it then fucked with it and broke it again 😂
Lol
ADHD got to me
Get a old xeon
It's an old Dell T7500 with the riser
aliexpress is easily the best place to get parts like this
Ye I could just get another motherboard, the bios chips are fucked on all of them though
My one had a custom chip that I put on and it was completely fine, but I wanted to take it off and mess with it for some reason can't even remember why
9TB of storage too
Need to allow links in here
Can't see
i intentionally prevented embedding because it's useless for this site
Oh
Well reason I say is cause if someone clicks wrong thing
Like discord scammers
Or fake steam urls
i assume not very fast cores
oh yea its x99 lol
Servers aren't for gaming though
wait x99 isn't dual socket, what cursed chinesium shit is this
probably 128?
Must be more surely
Up to 192GB on my old one
ok with those xeons, 256
Old server ram is so cheap
Was a misbuy
Sdimm I think
https://ebay.us/m/QOg9Up xeon cpu
Ollama
Talk about last minute! Use FreeTaxUSA if your income is simple
We turned vibe channel to off topic lol
saw this on twitter, not sure if it works as I dont have access to Opus myself to test
If true then instructions need updated
I managed to force HQ sky with LQ terrain (Terrain only renders in LQ) Just need to make it update with weather ⛈️
What the fuck
Ooh wait
Ollama does support cloud models
For a second I thought you were running it locally. I was concerned
😮💨 taxes done
Raytrace dem dolla bills
self employment tax is ridiculous
like yeah, i can't get hired anywhere so i scrape by with what i can get freelancing. but i have to pay twice as much because of it
🙏
https://github.com/Kim2091/Vibe-Reverse-Engineering/tree/renderdoc
can i please have testers for renderdoc DX9 integration? it should be supported on cursor, copilot, claude, and kiro now
I was using Amazon Q Pro with Kiro before, it worked like a full coding plan with all the Claude models, so I could run Opus 4.6, OpenClaw, and coding nonstop. But Amazon shut down my access without explanation, so now I’m using Copilot Pro+, Ollama, OpenCode Zen, Opencode Go and web model instead.
Is there a new renderdoc release to be paired with said testing?
yus
🤔 that's a good point
i need to add versioning to verify_install.py so when there's an update, it'll reinstall renderdoc

I use Ollama Cloud of course, my RTX 4090 can’t handle the big models. I recently bought a Mac Mini to run OpenClaw, but it feels too dumb even when I use Opus 4.6 as the main model. So now I just use the Mac Mini for training an ML model for stock / crypto TA.
For me, Ollama’s open source models are pretty good for most tasks. But if you want a cheap model specifically for coding, I actually recommend OpenCode Go, it works well with models like Minimax M2.7. For VRE though, it’s still not always good enough. If you have an OpenRouter account, you can try Qwen 3.6 Plus, it’s free to use for now. It seems to have coding skills close to Claude Opus 4.6, and in some benchmarks it even scores higher.
i was excited about GLM 5.1 mainly
😠 it's frustrating that it compacts convos at this usage amount
Y'all can try my branch that has a harness to use Gemma or qwen coder 2.5 for token savings
😭 are we gonna have conflicting PRs again
For me, GLM 5.1 is actually quite good, it handles many tasks very accurately. But for the VR mod, it keeps overthinking everything. Maybe it’s better for VRE, but I haven’t tested it yet. It only came out on Ollama a few days ago, so I haven’t had a chance to try it with VRE.
Nah it's not close to a pr state
that's really neat though
this is done now
@candid mirage this should fix the issues you had with renderdoc dx9 the other day
Actually, Kiro IDE/CLI is quite good for coding, especially for agent tasks. It saves more tokens than Claude Code, and its planning mode feels better in my experience. The agent hook is also really solid, I can let it code nonstop all day without any human interruption.
That's neat. What's the sub
Now to find a way to get Claude to use image recongition to test games on its own
Claude can read images, I just capture a screenshot and let it analyze it for VRE debugging together with the game control tools. You only need to save the image as a JPEG; no need to convert it to base64. The vision model can analyze the image directly.
that'd pretty rapidly waste context, but it'd be pretty simple
just package in a screenshot utility with vibe tools, and give it a good skill file on how to use it
Or just pass the image to Gemma 4 (local or cloud) for analysis.
i tried gemma 4 for this and it was very inaccurate when analyzing images 🙁
I use Qwen 3.5 2B GGUF (Q4_K_M) for image analysis through LM Studio, it’s really fast for quick comparison tasks. Things like checking whether a screenshot is black or not, or spotting the differences between before and after images.
It’s really not worth it now without the Amazon Q Pro plan ($19 per month), unless you can modify it to use a custom provider...
The AI is trying to gaslight me into believing there is no tracer option in the ffp proxy menu
Why I oughta...
The AI told you to reject the evidence of your eyes and ears. It was their final, most essential command
new smoking gun ✨
🧱 🦶 🔫
iirc Anthropic said no to helping the govt doing something (which is their right to refuse) but bc this administration is so blatantly corrupt they just get threatened into cooperation

Wtf man this abomination Claude generated. Kim new emoji for the server asap

Incredible
AGI finally achieved 
I could do better 

it does this every time
it just looks at the wrong folder, claims the exe vanished, then finds it and says it reappeared
gpt crapped itself like claude and drift off subject
damn does not give you much
i'm
done
i just
CANNOT handle it anymore
they nerfed EVERYTHING
everything is fucked up and not working right
ikr gpt was working this morning now its not
dont stress, its your birthday. relax a bit. it seems kiro qwen3 is doing its job on my end but thats a different sub module
managed to update some stuff before i ran out of free
anyone with claude subscription: use opus 4.5
it's 10x better than 4.6
it has smaller context but it's doing so much better that it just doesn't matter
Will definitely try that because... what the fuck is even going on anymore?
just tried going back to npm version 2.1.62, its back to doing what it used to
had this bug that it actually did like a full trace and root cause where new version itd try a surface fix and give up lol
something interesting i saw as well
i have the little inline context bar readout
and on new versions itd start at 0% even after first prompt
well yah know go up slowly
but on this old version it immediatly went to 6% after first prompt and started from there i assume it actually read the rules and claudemd cuz its follwing that better too
anyways def worth a try im using wsl claude code cli
🧑🍳
bruh
best part is i can probably get this to be single window with UI, man do i love java being easily decompilable 
im only doing this for beta 1.7.3 because thats my fav version and modern MC has radiance
ahhh lol
imagine seeing this back in the day lol
wonder how difficult it would be to get openmw working 🤔
yeah
remix api will be the best way in that case
just thinking of open source games I might be interested in working on lol
even with triangle backface culling off in remix, i think its doing some weird stuff with the alpha map
fixed, absolute peak

Is this actually Minecraft or another one of those like
Minecraft type apps
real beta 1.7.3
possibly? it relies on decompiling the game so def not a drag-n-drop mod like unityrtx
so you'd have to do RE work on the ver you want
i only want to do 1.7.3 so uh yea
Ah
Fair
Tu75 is all id truly want anyways
Or whatever the java version is
1.13 maybe
Thank you so much for this tip. It SO much faster and doesn't take as much usage. I feel like it is actually wasting less tokens than Sonnet 4.6 does.
can I beat minecraft but with rtx remix
this is the one version that you cant beat 😛 (well there is achievements)
damn
your sky looks better than radiance lol
atmos sky 
What does geometry hash even look like with Minecraft block based worlds
There's probably a block culling + large surface geometry triangulation to simplify larger structures. But I'm just guessing based off of a couple of videos I watched on how to make Minecraft clones in unity 
Gotta say though Claude can crunch through way more todos in a single prompt than GPT. I feel like GPT is "programmed" to constantly ask questions, sometimes ones that make no sense. It can be quite token guzzling
Trying to get a D3D9 renderer for dethrace (Carmageddon source port that's actually the full game instead off the more barebones OpenC1) going and it's definitely not going as smoothly as GTA 1 
If it could only find the smoking gun

It immediately went "wait no that's not it" but I'll take a smoking gun over a non-smoking one
I think Claude might be close to cracking FFP GPU skinning for OOT.
Not... everything lines up correctly just yet, but I assure you it all looks MUCH better than it did. The Remix toolkit even sees the bones!
That's awesome. Seems like it's coming along well
Honestly, this is one of the last somewhat major blockers, I think. There are some other more minor things like getting particles and clothing to tint correctly and whatnot, and making sure that all game-side lights are being passed through, but otherwise I think it is in good shape.
I will admit that I have not extensively tested it in a long session, so I cannot speak as to how stable it is.
haven't look at hashes but there is 0 culling
🙏 lemme know how the UI PR i did works if u end up using it
i will be trying it later today 
i haven't looked at it but i assume it's graphics api agnostic?
ya
it's not specific to UI either. you could technically forward anything from the original game
it's just that UI makes the most sense
🙏 dudes i am almost close to getting near instant cell loading for Skyrim and FO4
gm
wheres the stuck part
threading issues
if its performance, check for cpu bottlenecks with renderdoc. it fixes it
lul
HOW
did it auto export
might wana check drive space too
if it did that
my project is 10gb as whole
wait nvm half
5 gb
holy shit i did it i did it
it loads INSTANTLY now
no more standing around for 30+ seconds waiting for it to load cells
it happens as soon as you exit interiors now
nice
What was the issue?
turns out loading all cells and geometry on a single thread will slow things down quite a bit 😰
getting async to work though was a nightmare
it was using 50+ GB RAM for the last couple hours
Oh, you might've been IO-bound then. Sequentially loading world cells one at a time from disk 
If you were truly IO-bound, you should be able to spawn many more threads than there are CPU cores in the machine, since they'll be mostly "sleeping" waiting for disk reads
it doesn't seem like that was the issue actually. moreso... it's just a loooot of geometry to be loading
Debug View -> Primitive Index should tell you more or less how tris are distributed, if you're loading LODs correctly (I.e: not LOD0 for the entire terrain), once the geo is in memory, getting them into remix and then GPU should be really fast.
right now LODs aren't working at all. we are only loading LOD0, but the way it works is that only things that the game has set as LOD0 gets loaded. anything LOD1 or higher is just missing
i am afraid of how bad perf will be when we set up distant LODs...
whats the lod look like
might be an old bug we had for dx9 oblivion
remember theres ini files that skyrim creates
in documents
including terrain? I remember Morrowind terrain being split into chunks of like, 5x5 meters or so, and disabling culling to force it loading too far caused way too many unbatched instances at a time
second pic what is that? i had issues with a similar game
yeah it is loading a big square like that, but that's what the game itself provides. then i apply culling on top of that in the skse plugin to help with performance. but i can't really limit what the game is sending to the plugin directly
Geometry Hash
ah ok
because it's being done through the API, cell data is simply being captured from the game and then sent to the plugin. i don't think i can capture this info after the culling is performed
aw man im almost out of budget
Are you capturing actual rendering submissions or you loop terrain renderables yourself and submit to the API? Maybe there are better hook points that give you the final mesh to be rendered without having to do culling/optimizations yourself (aside from removing frustum culling). Should be more compatible with mods that change terrain rendering too
The plugin reads Skyrim's heightmap data (land->loadedData) and manually builds terrain meshes from it - extracting vertex positions, normals, UVs, and textures, then submitting the constructed geometry to Remix via CreateMesh. We're not intercepting Skyrim's actual terrain render calls.```
asked claude just to make sure lol
this method was used because it was what SKSE and F4SE provided
using the rendering submissions would remove a lot of complexity though, so we should probably just try to find the addresses for it huh
have you tried making a build without skse and f4se?
i can't do that
ah
Love it when the AI is like "should we continue or is this good enough?" when absolutely nothing works or has been implemented correctly
when you put a budget it puts it on the next bill right?
Oof, yeah tesselating a heightmap is nasty business. If you can find a way to just capture what the game intended to render it should save you a ton of time reinventing game logic + take advantage of the LOD system, modding compatibility, become faithfull to the game, etc.
did someone flip the switch on claude again
i have a weird issue where tabs flicker back and forth now
claude went to stupidville today
opus 4.5
it keeps trying to replace code thats already working repeadadly
You're not government, bub.
I wish I was government
I have run out of all my apis and subscriptions and I can no longer pay for another
Do they even have Claude access here or is just the US government
Gobbling up the resources
Probably us only
Claude for government def has seperate servers and doesnt get nearly as much usage as public, that shit will never go down
Claude Cowork will go down less, because what the hell is Cowork? 99% of users will just OpenClaw (of course Anthropic API cant be used anymore for it) or such
will give this a shot rn
ty for feedback
did you let nvidia know in discord
✨
e-man knows now 😆
Real question is, did you update your name on your taxes??? (jk)
humm
😭
glad you didn't change your pfp at the same time lol
one of my friends changes their pfp constantly and it always messes me up
1+ hour brainstorming session with opus 4.5 about skyrim's culling system and the correct way to intercept it
it was actually smart and not infuriating to talk with!
4.6 is so bad
also guys, you can use claude code CLI in vs code and it'll have access to your workspace and everything
it's def an improvement
still very easy to use
😔
im not testing it its too big
😔
imagine how many textures remix would have to export
after i get skyrim working and backport it all to FO4, i might try starfield just to suffer
starfield will use too much space with remix even . if the bridge does not crash
how tf is it so expensive still
they've updated space is free roam now
bridge isn't necessary, 64 bit
Please don't waste your time on Starfield
i dont rem buying this
it's the only game i put 100 hours into and only remember ~20 minutes worth of
hey @rose oasis Enderal special edition a steam skyrim se mod that comes with skse should be doable
nvm i forgot it has its own exe and stuff
😮 an ai pc
that $500 m4 mac mini sounds kinda good to use if this works
wait, how much ram does this need
good luck
Not sure, was just reading up on it, seeing it may be useful for some.
Ram costs are thankfully going down some.
what is swappable with a mac
ngl this looks like a very sloppy repo

would need more ram than luck here
how about a raspberry pi 5
the ONLY one i would consider running is GLM 5.1, and that requires an insane amount of RAM
i mention cause of it can use terminal instead of an os
apple isn't the only way to go here either guys. getting those AMD APUs with the 8060s is a more cost effective route
still linux, gui or not
the broadcom soc it uses is extremely slow compared to even a macbook neo
ah
fair enough
call me when they actually invest time into their compute backends
rocm is the biggest PoS i have ever used
on linux it's actually very good already for LLMs
no toolkit for linux yet right?
that was not my experience with the 9070
so many random GPU lockups trying to do any inference
not to mention how shaky the support is in general with popular tools
hey @rose oasis can you make a linux branch of toolkit for linux with ai?
they used vulkan inference
yeah vulkan worked well for me
you tryin to kill me lol
😔
The one dude I know who uses multiple 5090s to run glm
qwen 3 is ok but it has its own issues with coding rn
using multiple arc b50 or a770 gpus is prob another more cost effective way
i had to use claude opus 4.5 to look at the build errors
arc b50 has 16 GB VRAM and is $350
i have a intel arc b580 oc 12 gb edition
GLM 5.1 is a MASSIVE local ai model that you might be able to run. We are checking out the performance with llama.cpp and unsloth quants in this video.
Local Ai Software Setup Guides https://www.youtube.com/playlist?list=PLarJAzZsWRGA1Ej0Kxt1z6LAtgtCUWMtO
Written Guides https://digitalspaceport.com/how-to-setup-an-ai-server-homelab-beginners-gu...
lol
how is it
very good actually
although it has its own issues with remix with vulkan
i found ways around it though
Unfortunately the 32gb B70 is just too expensive
but still hedious
I think it goes for around a thousand
yeah
but then compare it to the only other 32 GB consumer GPU in existence
5090, $3k
Can you game on it 👀
you can
games run pretty well on intel arc
Ice cream?
yum
yum
i'm having hard apple cider
Here
1 hour in
without zooming in i saw potatos with gravy
Is this a dog bowl
Metal human bowl
That said. Is there really a difference?
make the bowl bigger
😔 it's 128 GB RAM max

✨
i have 32x2
remember when people went crazy over apple overcharging for ram
good times
apple somehow manages to extra overcharge now
only thing i rem about apple is that you couldnt modify the pc
honeslty though glm is good enough that it may be worthwhile at some point for some to invest in the RAM to run it locally
no more $200 subscriptions lol
it gets pretty damn close to Opus 4.6 in coding
Glm is very good
If you ever get it running. Let me use it for inference
okey
Big doubt
¯_(ツ)_/¯
where is the og one
I've taken a couple different things that I probably shouldn't have.
But at least I'm not addicted
Ezfn (dw mods epic allows it as long as they dont add purchases)
Youll have to do a tutorial online for it
Its been to long since ive installed
https://huggingface.co/inferencerlabs/GLM-5.1-MLX-2.5bit-INF
https://www.youtube.com/watch?v=Sx8IPdbMlok
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Ever wanted to run the 1.5 Terrabyte GLM 5.1? Well soon you will be using extreme quantisation techniques shown in this video.
TEST SYSTEM
Inferencer App: https://inferencer.com
2025 M3 Ultra Mac Studio | 512GB RAM
Q2.5: https://huggingface.co/inferencerlabs/GLM-5.1-MLX-2.5bit-INF
BUY NOW
Mac Studio: https://vtudio.com/a/?a=mac+studio
MacBook ...
wow
i tried things like this with opus 4.6 and even that failed
There's also REAP. I was able to run glm 4.6 locally with it.
It was slow, yes.
But it worked
😭 why tf am i considering this
how much does their ram cost now?
it would take over a year of me paying for the $200 claude sub for this to even get CLOSE to paying itself off
and i won't be doin that
Idk if the reap team has worked on 5.1
They used to sell soddim chips at a 10x markup few years back right?
it's $6k for their 256 gb mac
what its this chanell?
Yes that's "normalish"
No other agents sorry thats baned
most of them are out of stock, even 64gb ones
Off topic 2
people are catching wind... 🙁
wish i didnt sell my m4 pro mbp
😔
there are better options anyway imo
esp at those prices
6 way Arc b70 is gonna run faster and be the same price. plus you can do a lot of ohter stuff with it
@dull swallow
though that is only 192 GB
Dang
Trying to get a D3D game from 1997 to path trace is frying my brain
what game? lol
damn
which one
blood? blood 2?
etc
MDK
At this point I'm not even expecting great results. I just want to try to have any in game polys show up path traced
i wonder if decent can be looked at
cool kids club
If there is one thing I wanna get outta this is being good at vibe reverse engineering path tracing into old games
Logically there is nothing preventing it
why not port them to native windows
we do have re tools
guess game
still has ui issues
i cant select checkboxes or settings some reason
it will check and then disable
game was Mortal Shell
😠 and it makes me tired
that's all it does to me
i sweat and i get tired
😆
I really don't understand how an AI can be so dumb to rename a dll and then be surprised that a new log file written now has that name but at the same time produce any usable code ever
wait r u trying to launch an old build of fortnite thru the main exe?
oh nvm ur using ezfn
wait ur using ezfn?
Why not use a different tool that lets you play that build offline (lawin server, rift, era)
i didnt know there was an offline build
there’ve been tools since 2020 to play fortnite builds offline
installing fortnite again. good thing i have 2gb up and down
Carmageddon is the loudest game ever and the AI keeps opening new builds in the background wrecking my ears
Love the phrases we come up with
Silent Black Triangle
Silent Black Triangle vs Loud White Circle who wins
those are no longer being worked on. i tried everything
even the domains are gone
you should try carbon then, sadly its not open source so use it at ur risk
not touching it then
we have a thing on this discord about non legal stuff
we dont do it
im only using my disc 1 from my dump until i get that working right
DuckStation
So who's going to use this job stealing, planet killing AI to retopologize 3d scans for me with a fancy new blender plugin?
It's such a robotic job to do, you'll all be rich if it works
Same for UV unwrapping
AI UV Unwrapping when
(skill issue on my part tbh)
Oh mate a man can dream though
AI Selection tool while we're at it 
offtopic 2
Sorry, have at it
peak
peak
btw how's this going
@bright erminecan you tell me what you promted and used for minecrafT?
with a capital t
gpt 5.4
ok
ive figured it out, #offtopic is secretly #wordle and this is the real #offtopic

off-topic-4
What if the real #offtopic was just the friends we made along the way?
pretty upsetting tbh
yehahh

i wanna try using my credits for minecraft tu75 but we dont know how to do any of this properly still
and itd prolly waste too much
to learn, i'd recommend doing a game that's already been done. like outrun 2006. you'll learn a lot from that, and have a good reference to compare with if you get stuck
def hit the frustration detector with that one...
reasonable
i mean minecraft was done by cr
combined with finding out that the only place i found to rent wants me to earn 9k per month... 
but only for beta
yeah i meant to learn how to use these tools though. newer MC could probably be done still
ah
😠😠😠😠 i'm havin a bad night
basically all of my moving plans blew up 😭
aw noo
sorry to hear that that sucks
for the fucks of it im gonna barley try wolfenstein the new order
to see if i can even get anything out of that
would you recomend a to dx9 or remix api for that
cuz its uh
pretty far opengl
api
kk
so id start with prompting it to trace shader calls/geometry data and decompile them?
sorry didnt mean for 2
2 always sounds mean
unfortunately i don't really have good guidelines/recommendations
so far, everyone working with the API for these use-cases has either had access to the source code for the game/engine, extensive documentation on how it works, or a modding toolkit that has most of the info required
this game does not. you're prooobably not gonna get very far
no ur oki
tno would be cool but ion wanna waste credits
yus 😖
a great possibly?

hm
everything we have is either unity or closed source or dx11 bleck
welll

my only thing id actually try with no real like
source or whatever is psychonauts 2
that one miighttt work
like
if the game has clear strings that the LLM can search for annad figure out, it might work
https://github.com/bgbennyboy/Psychonauts-Explorer i mean this does exist idk how useful itd be
wait thats 1
nvm
our only only fear
idk if this is good or not
ik ue has common strings (at least for 2.5)
yeahhh
so for that case i would ask it to look for common ue4 strings related to geometry and textures first?
yeah
downloading that massive list of common console commands and giving it to claude or w/e to search though would prob be useful too
ya

"using the workspace (idk how to point to e mans decomp or if i should point to something else?) as well as UE4console commands try to find and locate common strings for rendering with this game for geometry and textures"
or should i say more?
first, you want to clone this locally: https://github.com/sambow23/dxvk-remix-gmod/tree/unity
- dxvk-remix-gmod - this is a customized version of remix that has UI and skinning support through the Remix API
- game folder
- project folder
within the project folder, i have UE4consolecommands.txt. you can grep this to find potentially useful strings to search for in the game files/binary. ask questions if you are unsure, do not assume the answer```


hopefully today will be a good day though

