#Vibe Reverse Engineering Toolset
15844 messages ยท Page 16 of 16 (latest)
I see
just mostly fixing stuff in the custom remix
I usually use claude and gemini, I geniunely don't know how flash sometimes solves issues that claude doesnt manage to see
Okay
It's a new day
Yesterday was thoroughly painful
Let's see if shit works
lemme get my cough syrup first
get well soon
The best one that you can actually run at home is Qwen3.6 27B
From what I've heard it is actually somewhat capable
But requires a lot of manual context management and shit that is usually handled by your average closed source model
Thanks for the reply! Shame, I was hoping some sort of probe to sample lighting information on the character would be possible. For the manual path I already had an idea about volumes that "eat" the lighting, already set that up in a few spots. I guess I'll just try to use AI to help set those up throughout the game.
I'm on my adding stuff to Manhunt bullshit again even though I told myself yesterday was a good stopping point for an update 
@crystal kiln Sorry for pestering you again, feel free to reply later or when you want to but you're the particle man so I thought you'd be the most qualified to answer. 
I have a hard time wrapping my head around the use of RTX Remix particles or the best practices on integrating them into games.
Let's say a game a has a old looking blood effect I wanna replace. It generates a bunch of puffs and blood particles, each as their own draw, with a certain logic of how it behaves.
What would be the best thing to do in this scenario? Replace the individual effects with single emitters instead and then set up the RTX Remix particles around those? This would require some modification of how the game generates the particles in the first place, right?
Hey, no worries. Feel free to ping me whenever you need, even for simple stuff. Happy to help ๐
So Remix particles were designed to be used like any other replacement asset. You can target materials and mesh hashes. When remix particles are attached to materials, all meshes that use that mat_hash will start to emit particles from the mesh surface according to the settings you chose. When attached to meshes, then only that mesh will emit particles.
Since we're talking about dynamic meshes without stable mesh hashes, you'll prolly want to promote the blood particles mat_hash to be a particle emitter.
The thing is, you'll need to author remastered textures with airborne blood splashes/dropplets, probably as a sprite sheet, then configure the mat_hash with the sheet rows, columns and fps.
Then the remix particle system allow you to set the sprite sheet mode to either be animated, or for each particle to pick a random sprite from the sprite sheet. Whatever suits best your visuals.
What are you most unsure about this process?
I guess for the games I looked at it's always the issue of it being a bunch of particles spawned with the same texture, so targeting a material hash as a base emitter wouldn't work, since the same material appears multiple times. And the individual spawning sprite meshes have unstable hashes, plus in some cases like Manhunt, also spawn multiple times.
Like in the case of Manhunt here, it's three puffs and a also a couple of spray meshes that make up the effect
All using palletized textures too so targeting textures is also an issue on that front
I don't get it. Does it appear multiple times for different purposes, not just blood? If so then yes, these particles are too generic to be made blood-specific.
You'd need to mod the game itself to spawn a unique texture instead
Yeah, they are reused for different purposes. I've come across this a bunch for games of this era. They love to reuse textures for smoke, fire etc too. I guess modding the game is inevitable in those instances. So a single mesh/texture to attach all Remix particle logic to would probably be the best? And then handle the spread etc just through Remix?
Depends on what you want to remaster. If it is just the blood particle, you'll need to explore a way to decouple the particles from using the shared textures for a dedicated one and replace it. If you want to enhance all particles in the game you'll want to enhance the generic cloud puffs and what not.
It really depends on how the game uses and how the textures are arranged.
I never played Manhunt so I can't give too much advice without looking closer in the textures and draw calls, unfortunately, but shouldn't be too far from this
Yeah, I've been working on that today, trying to bake out the colors rather than having the shared particles. I'll probably do all that work first and then get back to the particle system later. Thanks again for the help!
Thought one of my changes broke something but then I thought "wait let me google that real quick" turns out it was just a game issue 
The Wire or the Machete can be spawned in this scene via trainers. However, using them to execute hunters will crash the game. This happens because the severed head is missing from the scene files.
Tried testing the Machete in the level with the quickest time to gameplay, turns out you can't do that because it causes a crash
Also looking at blood sprites and executions is a weird feel let me tell you
Ah right, a crash when trying to behead a headless creature 
why does ai keep wanting my project to load dlls from system32 when i tell it game dir only
AI or the game
The way DLL hijacking in Windows works is that when not specified, an executable first looks for the DLL in the same folder it is located in. So when the app calls a function, your fake DLL runs its' own code first and then forwards the request to the real System32 DLL
So you're still using legitimate system DLLs
If the game uses hardcoded absolute paths to the real thing then you need to patch it somehow
yea i can see that but its like it sees the proxy i made and then it just skips it all together
falling back to system 32
my proxy loaded a direct x 9 dll with the proxy at first. then it did something somehow and now it does not see the proxy or the 9 dll at all
i even tried messing with the initializer but nope
misaligned and trying to hijack your OS. bet.
jokes aside im serious
Okay, Claude lowkey sucks
It assumes cornerstone info that it has direct access to, instead of just... Checking
this is the biggest problem with 4.8
I've been using 4.7 more
i spawn 4.7 sub agents to check its work
I like 4.8 but it's a bit crazy. and paranoid imo Which paid off though because it caught 7 out of 8 subagents it spun up hallucinating and assuming stuff in their own right. even before it spun the first one up I seen in it's thoughts that it was mistrusting of the subagent from the get go. after the first one it gave the subagents explicit instructions that they all choose to ignore except the last one.
Holy
Bloody
Meme
So
When Codex is out of limit - it finishes task and you wait for restock
When Claude is out of limit - it fucking STOPS
So it just borked me shader and render script, and instead of finishing work with them it just said "phew, I'm done, see you tomorrow"
47....
Stellar
47 is no longer the star trek number
am I old ? ๐ด ๐ด ๐ชฆ ๐ชฆ ๐ชฆ ๐ชฆ
Yes
What the hell is bro reviewing
Headlights for all
I'm doing all this work for Carma when there is still the issue of the game gobbling up VRAM unless I boot it with RT off and manually toggle it on 
I managed to get it working I guess but I geniunely don't know which version/render I should be using, there's two right now that I can get to if I remove certain onscreen textures
They both have working RTX remix features, just that one is buggy and the other one isnt, the second looks alot more realistic and the first isn't all that
second:
oh right, here's the original version for comparison
this looks like a flat rendertarget
the second one looks like actual remix
Yeah I guess you were right, the AI kept telling me it was the first but I don't really think so, the second version has wayy more stuff that the first doesnt by a good margin
although with more bugs than i could ever imagine
would make sense too yea cause
it may not fully work ever period but its not not worth tryinhg
that
I mean, I'm willing to put in effort to see what I could take
yeah
Sparkles disappears in a poof of pink sparkles. Fitting, I suppose.
Anyone know if the newer dxvk remix releases/commits handle spatialmap:insert duplication cases?
I've been tracking an issue where certain models of the same kind would try to connect themselves to eachother when I'm in the raytraced view, and when I checked the dxvk log I found
"[11:34:59.967] warn: Specified hash was already present in SpatialMap::insert(). May indicate a duplicated overlapping object."
or is it probably on my end to fix?
Carmageddon low key be looking atmospheric like this
I fixed the VRAM eating issue by just putting a delay on the FFP conversion at the start of races 
If it works it works
I've ohnestly gave up on trying to fix that issue cause it drove me in circles but nice work
Yours was only when looking in certain directions right?
no like just generally turning it on without the Z threshold set to X amount would do that
it gobbled up like 8GB of vram in 10 seconds ish
probably yeah
What game are you working on?
me or gabdeg?
Oh sorry for the confusion, I was talking about what you're working on
oh, no problem lol
I'm working on payday 2
its a shader model 3 d3d9, I was told it was unlikely but like I figured its worth a shot
so far its promising but like I'm a bit stuck on the models of the same type merging
Hmm, it's on diesel engine 2
Yeah, janky engine
The problem might be inherent to the engine given what I've just read about it
Was initially built for a driving game by GRIN
But that shouldn't really be an issue
pretty much
even claude gave up ๐
"At this point I want to be honest with you โ we've exhausted most approaches:
Matrix injection โ
Reg0/reg6 world transforms โ
Vertex buffer position extraction โ
rtx.conf hash rules โ
uniqueObjectDistance โ
fusedWorldViewMode changes โ
Register 255 picking value โ"
bro pulled out the fancy โ
I'm guessing its a collateral tbh between diesel and how duped meshes are handled or whatever
no idea why lol
Did you not ask it to do an api trace?
about that lol, I'm still using conversation based tbh, figured I wouldnt need it but now I'[m really considering it
is there a level editor?
On pcgw I'm seeing some mod managers but I'm not sure if they allow rendering level access
as in like this
custom ones for sure I saw ones floating around
Oh so like no ida tools
pretty much
Celer looks decent but it's doing the opposite for remix purposes
but I already know they get fully merged cause in the captured scene, and even in the live texture editor, when I press on the cone it highlights the whole three
You really should get an ida tools and ask it to make a tracer for the games api
Pulling out the big knives too early
What
You gotsta use the toolset with actual Claude Code/Codex/Windsurf or whatever
Payday 2 is too new. It needs serious work
Yeah
@sharp seal funny texture
wait.... @golden raven do you know about this? https://github.com/SpeedReflect/Binary
Yeah hold on lemme mess with this tool. Seems to offer some game tinkering aspects other than textures
My god. Finally
So what good is this for you after all the AI yapping
Thank you very much and to the other fellas that helped, I managed to install one of them and I suppose it applied the fix since I gave it context and the specific line of code
the build dxvk is very annoying
its so buggy
Yuck
Claude, please never use memory
He probably doesnt, but i do. Its literally the basic modding tool for underground 2 to change textures and functions
Carbon, MostWanted, ProStreet and Undercover than swapped to vlt data.
I read their readme file and it says you gotta put a password to use the modder tools (from their discord ๐)
Just use this: https://nfsmods.xyz/mod/1638
Its version 2.8.3
(dm'd)
What game is psy2?? ๐
Psychonauts 2
Ooooh very nice
how does british guy got the same time zone as me
maybe I don't live there anymore M8
i always thought all brits lived in britian
no ones meant to know its a secret but I trust you
Fuck, I'm thinking of getting out of the AI game for a while and a Armored Core 4 recomp is announced
Coming soon:
Armored Core ReAnswered:
- Static recompilation (native on PC, 0 emulation)
- Via Reglue SDK & Vulkan renderer
- Huge performance improvements
- High FPS, PC native multiplayer
249
548
You know this would go hard with path trace
My GPT sub ends on June 5th, Im also contemplating getting out, already like $400 in over 3 months. I dont really care about the cost, Ill only ever have 1 sub going forward, but still. Shit dont produce results by the 5th, then Im all aboard the AI hate train
bro.....
You stole my name idea!!!!! 
It was, in fact, not.
Seems like the Armored Core recomp is based on Xenia so probably not an easy job to create a new fixed function renderer
Xenia as in the sonic recomp thingy?
xenia is the 360 emulator
there is a recompilation tool that uses the Xenia JIT. this is good in terms of compatibility and how easy it is to use, but it also comes with the drawbacks of Xenia
i guess it captured the skybox mesh?
o.o neat
no normal geometry, and my screen now resets after a bit of time
we dunno why
hmmmmmmmmmmm
That's cool
How did you grab the cam
Since it's dx12 translation of geo and textures shaders and all of that is rough
Games can run in dx11 too with the arg line
Check the textcoords and the world matrix as well?
Could be rendering the wrong cam too
I had these issues making a remix dx11 project that I have in discord
wont the excessive yapping take more tokens?
โ
It will but ai tends to gaslight
My GOD this was 3 days of reverse engineering.
@crystal kiln I need YOUR help. For you know. The project that you joined and are morally not allowed to leave until you see it through.
What is the best way to create dynamic lights via the api? Currently, they're pretty unstable.
They are quite stable, something is off with the way it is capturing the car position and light tubes position every frame in the mod, also you're creating lights with different hashes every frame which doesn't allow denoisers to accumulate light info across frames.
Try reusing the same light hash computed on the tube shape and color, and find a way to stabilize that jittery anchoring
Do you have any referance code for this?
I've never worked with api lights before.
Not exactly, as this will be something in your mod code to change, not specific to the remix API usage.
The hash is something you feed yourself in the LightInfo struct when calling remixapi_CreateLight, and the lights position/shape are derived from your code in the mod
https://github.com/NVIDIAGameWorks/dxvk-remix/blob/main/public/include/remix/remix.h#L995
I assume you're already using the API, no? Should have already setup the LightInfo + pNext for cylinder lights by then
Right now, the lights never survive past one frame. They get deleted as soon as they're created. I can spam force recreate and watch them exist for one entire frame over and over again.
I also summon @violet flume . Trying to work with API lights here. It's been a real headache.
@crystal kiln
Remix Plus has some weird api light persistence code in it that just wouldn't work for me in IV so I got rid of it completely ๐ I might just not know how to use it but I couldnt bother reading the AI written docs
No way for looping code? Jumping to the start after creation?
try with og remix until you get them working correctly. As for the jitter, might happen because remix sometimes uses your position data of the last frame and sometimes the one from the current frame or something along these lines? When are you creating and sending the data to remix? Begin / end of a frame?
wouldnt worry about the debug view too much tho
When you create the light after you destroy the one frame then create a frame after creation. Then have it loop back to creation therefore creating a loop
And the hashes?
mmmmmmmm
It will stay on creation therefore rest is null
I have a feeling that the light position view is fucking lying
The hashes are unstable but fine.
unstable hashes work just fine
But the light position gizmos aren't really showing the truue positions
they are in fact unstable
do you want fast moving lights to accumulate?
Make it a single light?
would leave somewhat of a trail behind you and you move so fast that there is nothing to accumulate anyway
Can it be faked framed?
I suppose you're right but remix could technically track the light motion to kill old rays and create new ones ahead of time
idk I'm talking out my ass rn
I'm doing my best to track and re-use existing lights to keep the hash stable but I have no way (not enough unique data) to uniquely track dynamic lights
Can you look via hex?
headlights get quite noisy when you are moving fast, increasing rtxdi sample count is the only thing that helps
Ah now I see you're using the bridge path, not the C API. Via the C API, you would just call CreateMesh with the same LightInfo with update positions and same hash, without calling DestroyMesh.
This would allow the light to survive and accumulate RTXDI buffers for denoising, light picking during sampling, use temporal data and so on
having an option to prioritize lights would nice
Do I smell an possible implementation
To the ui
K
When I try to help w/e
In theory restir reservoirs should do that by finding best nearby lights to sample according to what neighbor pixels found with random sampling. Not sure if biasing sampling towards a few lights will truly yield better results unless you crank up sampling count like you did
Having much less lights in the scene improves headlight noise. Gets worse the more lights are in the scene
That is expected up to some extent as they're all fighting for sample budget, and gets worse if these other lights also have a considerable importance. The more you bias towards one light, the less you sample the others, so to not leave anyone behind you kinda need to give each light a minimum amount of samples or they get eaten away
Mhh true. I think the default sample count is like 3 so with 60, having bias towards 2-3 lights that are always sampled shouldn't hurt ๐
But I might be completely wrong on that (just talking out of my ass as burrito would say)
Nah it is good intuition. The whole balancing act of light sampling and picking is quite tricky
Tbh I really wonder how big of an overhead increasing SPP would bring today. It definitely won't be plain x2, but it won't be slim either
The key point I guess is that the current SPP already works well for the majority of different scenarios, and increasing it will unnecessarily add overhead to scenarios that are working fine, only to help with extreme scenarios that can be noisy/inefficient for many varying reasons, not just sampling amount but quality too, so the goal now moves from working harder to working smarter -> Identifying the extreme scenarios (and why they're bad) and adaptively perform samples or techniques to clear the artifacts.
Sure, but there are still scenarios where more SPP will help much more than any current approach) Plus stuff like better handling of upscalers ect. I am definitely not saying that current approaches should be scrapped in favor of bruteforcing, but I do hope to see some increase in SPP someday
Also, that might be silly, but how much of fine control can we have over sampling steps themselves?
Just as a surface idea, but having multiple NEEs on first hit would help with those specific situations like fast moving headlights, as well as stuf like shadows ect
Is it not realistically feasible or just will bring less benefit than headache?
Also that universal overhead always made me worder about of possibility of "path tracing VRS" so to say - deliberately spending more samples on parts of the screen that need it more, while leaving less complex scenearios running leaner
True. If you can afford to bruteforce to overcome an artifact then increasing SPP is necessary, only because we still don't have a better smart approach for that artifact since it is still an unsolved problem (or just not implemented yet). Think caustics for example.
But the goal should be, as much as possible: To keep minimum SPP covering most of scenarios and work with smarter adaptive sampling for the edge cases, otherwise the renderer gets lazy and unoptimized.
E-man, what do you think of light portals in RTX Remix? DS2 Lighting Engine has been implementing something like this and I tried to put something similar in RTX Remix. Is it silly to do because RTX Remix already does robust importance sampling?
#1481706093976621277 message
Here's how the author of DS2 Lighting Engine describes it:
Yeah extra shadow rays can be expensive, to the point that whenever we need to solve a problem that would require adding a new rayquery, it is highly profiled with stress test scenarios and usually gated by an RtxOption/Alt+X menu tick/slider
You don't. THat stuff is just kinda not simple)
Was in special education in high school
That is the right thinking, working smarter. This comes with drawbacks ofc, parts of your frame can be undersampled, have temporal artifacts and what not depending on the technique employed, so the great question is: How to identify when to and when to NOT sample.
Think of, idk, checkerboard rendering where half the screen renders at a time, half waits for the next frame, in a checkerboard pattern. Half the pixels will be delayed
Oof. My apologies!
It's ok it had to come out sometime
Interesting. Is NEE and shadowray still different things in full PT?
Path tracing calculates from camera position to the light position.
If you have bounce lighting entering through a window (indirectly), there will a lot of noise.
So to improve the noise levels in the room, you put a light at the window and tell it that: You won't emit anything but you will guide the rays towards yourself.
This makes sure that the light rays have a direct path to follow from the camera to the window.
That decreases noise.
Or it's more like shadowray being a part of full NEE estimation?
Yup, that always what made me HMMM in that idea - how to actually filter VRS rates across the screen
So just the rays and no emitter?
hmmm interesting. I don't see why it is not achievable by slapping a rect light though 
Some type of local convergence metric?..
Idea is - door itself doesn't emit anything
Portal just gives integrator a hint that there's a lot of tasty light BEHIND the door
Correct. It will guide the light rays out the window and then they'll hit the skybox or anything else. It's simply a way to tell the ray where to look for the light because remember. In PT, a ray is shot from the camera and it's told to find a light. If it does, the pixel is given a brightness level.
But how can it find light besides the sun
You could also use emissive but the idea is to use the natural lighting of the scene
It looks for an light out the window. It could be the sun, any other light type, bounce light or simply the skybox. The hardest path is for the light is to actually find a point of escape out the room. Once it has, it can easily find a source. The portal simply helps the ray leave the room.
Especially in games like Dark Souls where it's just sky light and no light source indoors
a shadow ray is just a ray that is cast from a point towards a light to see if the light is occluded by another object (shadowed).
NEE afaiu is more about the recursive algorithm where multi-bounce lights are broken down into recursively sampling lights directly in each bounce instead of doing totally random hemisphere sampling expecting to hit a light
https://www.youtube.com/watch?v=FU1dbi827LY&t=8m36s (see at 8m 36s)
Shadow is simply the absence of light.
Beautiful.
So no emitter and the shadows stay unless casted upon by light
correct. Shadow is not a physical thing. A shadow just means that there's no light there.
In perfect path tracer no light = compltele black
No "glow" that we get in raster graphics
That's why PT graphics can look much brighter (a lot of bounced lighting) than raster AND a lot darker (no real light reaching the area) in different areas
if only it were that easy
So just rays and no light source
Well, no, there are light sources)
Just light from those sources makes things lighter only when it can actually reach them
While in raster you have lights, sure
So light bends when reaching?
But you also have, so to say, "everpresent lighting" which illuminates everything, even things in completely covered areas
Not bends, rays are always straight. But it bounces, and bounces A LOT
What if you are in a box with no source
Yup
In PT it will be completely black
In raster it will be "dimly lit" in most cases
Because raster can't bounce lighting
And without that dim light shining on everything shadows in raster will be completely black
Well, sometimes they actually are - Doom 3 for example
Can you cast a source from cam
Yup, it will be essentially a flashlight
Ah
Explains why a lot of games have one light source
And if you multiply it it will get brighter
How do emissives play
I'm sure textures would be dark with light
Not really. That always present glow is separated from the light sources in the scene, so even if there are many lights, areas in shadows will still have same amount of light. Problem is - even when you have NO lights - those areas (and everything else) will still have same amount of light, instead of getting very dark
Games solve that by basically manually tweaking that everpresent light based on time of day, player being inside/outside ect
Ahh with @golden raven explanation that makes sense. It is not for artistically adding lights, it is meant to guide GI sampling direction. Interesting. I wonder if it helps against Restir GI & NRC because in theory they're already generalized guides for such things. Might need to check some A/Bs with DLSS off
Alright. Fiiiiiiiiiiiiiiinally
It is done. Yaw, pitch and roll all accounted for.
Lights stable and follow the car correctly.
Gonna fuck with trunk and engine neons now. See if I can get them to work.
Should be easy.
Technically they aren't visible during normal gamepla but I'm curious
Trunk+engine neons working.
Is it all hooked in, car models, colors, blinking animations, etc.?
That's next on the plan.
You have no idea how long I had to work on this.
did you ever come up with a good solution for clear coat?
It certainly increases sample in the area, but the way I implemented it it still looks like an emissive light kinda. Noticable effect near the portal but the light/rays aren't guided further in.
In DS2 it really brightens up rooms which is what I was aiming for. Couldn't quite figure out what to prompt to make that happen.
Here is what the author of DS2 Lighting Engine was able to do with them
(Not RTX Remix)
basically helps indoor environments to move more closely to reference?
Like months and months ago, yeah
It might make more sense to them if they're not using restir gi or rtxdi. It requires more work though since you have to go and place the portals
This video is a mood. I don't know what mood. But it is.
he's using brdf sampling + importance sampling for direct and sharc for indirect iirc
I wasn't sure if you scrapped it or not lol
whats brdf sampling
honestly google will be able to give you a better explanation of it than I can lol
this is why i asked
so it renders colour/textures better?
Don't shoot rays randomly in all directions, pick better directions that will make the lighting converge faster/are more likely to have meaningful information.
It is a whole family of methods to make smarter decisions about which directions to cast rays towards.
Light Sampling for example casts a ray towards lights instead of uniformly random in the expectation of one ray hit the light.
Yeah, that's why I was asking if it's a silly idea to implement in RTX Remix since it's more advanced anyhow.
ah so one of the methods would be importance sampling?
It is an interesting concept, if it works that well. Remix already have portals so in theory you could place one portal facing forward and the other backwards in the same door opening. Not sure how Remix treats portals regarding sampling guiding but I assume it would work. Is that what you did more or less?
Hmm, I only placed one portal pointing inward cause I think that's what was done in DS2 lighting mod. Should another pointing outward help out?
The AI definitely implemented it based on stuff present in source code, although it could've just been gaslighting me.
I think it was this
- Importance Sampling is a general term from statistics/Monte Carlo problem solving for picking values that better match the random problem you have at hand with educated choices rather than naively sampling it randomly.
- BRDF Sampling is one technique used in importance sampling where it is "more worth it" to sample shiny objects towards directions where it is more likely to reflect vs diffuse objects who receive and scatter lights from/to multiple directions (The surface BRDF function sorta instructs which ray directions contribute with most important stuff). Good for GI/indirect lighting.
- Light Sampling sample directly towards the scene lights with the so called "Shadow Rays" instead of randomly stumbling on a light source by chance. Good for direct lighting.
- Multi Importance Sampling is when you combine several of these techniques (i.e: BRDF & Light sampling) to get the best of both worlds, assigning weights to them so they merge/blend their contributions and what not.
and im guessing remix does Multi Importance Sampling
Here is a couple great references:
https://www.youtube.com/watch?v=p772XkEnEIU
https://www.youtube.com/watch?v=0IrzX4LDIx8 (at 31:50 onwards)
Is this gonna make limits better for us or worse
better for 2 weeks, then much much worse

They are probably going to make it more expensive/shittier to please shareholders
Maybe I shouldn't take an AI break and get one last month of usable AI before it becomes completely unaffordable
just one more month
lmfao real
i pray local llms catch up to a semi-usable state
me too
๐ฏ it's fucking over, and whatever little bit of good Anthrophic had in terms of morals or whatever is also gone after this
Meowning (meowvening) everyone
Lookie, actual shapes
Turns out my Copilot usage is actually still going till June 13th, so AI break isn't happening yet. Lets see how many credits I'm spending on a single prompt
New weirdness unlocked, RR makes the lines on the road look blue. Anyone know what this could be?
Also no idea why the video is like 5 frames in total
Hmm, one prompt, 1% used. Not great but also not terrible
alpha blending did that for me a long time ago
with oblivion
ok
That'll be $1 please
Seems like Copilot Pro+ is now roughly 1 prompt = 1% of your monthly usage (Sonnet or GPT 5.4)
back to square one for real
has anyone tried the fake camera injection github repo?
Well it was made on this very Discord server๐
https://discord.com/channels/1028444667789967381/1423046815112101938
are you using the fixed function proxy for your project?
if so it has a mode to force the fake camera
Yeah Iw as considering using it but I wasnt too sure
Sorry, how do I know? what I know for sure is that the game uses non fixed functions
Cause take a look at this, there's approximately 44 cameras, one of which has actual coordinate, I manage to force use it via AI, but the camera's position or direction does not correlate to the data I saw in the file
You can use it for funsies to see how far you can get with the fake math. Other than that, it's not really useful and I don't think it was used in any serious project
so, VRET comes with a fixed function conversion tool that i made. it can convert games that use shaders to a fixed function pipeline. if you ask it to do this, it will use that code
if it does use it successfully, you can press F4 to open the menu in game, and turn on that hack
I was originally gonna use it but like this gave me some sort of hope, weird enough Remix doesnt seem to be even bother to update its location in the UI
[PS] Camera 1: Reg 0 (view at 4)
Proj: 1.4000 0.0000 0.0000 0.0000 | -0.7722 0.6352 -0.0122 -1121.5621 | 0.0000 0.0192 0.9998 140.8086 | 2.0000 1.0000 1.0000 0.0000
View: 19200.0000 19200.0000 2400.0000 0.0000 | 0.1700 0.0000 0.0000 0.0000 | 0.6800 0.0000 0.0000 0.0000 | 0.2833 0.2833 0.3500 0.0000
That's pretty cool, and there wouldnt be any drawbacks?
Depends on the game. If you are very successful with the conversion, the only universal drawback would probably be CPU bottlenecks I think
I'll give it a shot soon then, thanks alot
I'm not even sure if the camera is the source of my issue tbh, it could be, but its the logical thing I could come up with, cause the agent tried multiple times to edit the SRC of the .trex d3d9 to block merging but that all wasnt of use
even in games that do provide fixed function fallback, this rendering path is always much less performant
cause I can see all the rendered raytracing, but models, (certain models to be exact) try to connect together for some reason
(ignore the camera, i was modifying its positions)
my best guess was that it was cause the camera I had wasnt properly updating its positions etc
the drawback pseudo mentioned, plus that some games might not play nicely. even if you get the conversion to output valid graphics, there may be a lot of specific issues with certain shaders
for example, in FNV the water is missing. so it needs a custom patch. the LODs needed custom handling too. the terrain also did
I see, much appreciated
Yeah. Also note that even the official Fixed Function implementations can be wacky and lack some important things. In Painkiller for example one of the "signature move" effects is missing because it required dudv texture addressing not available in DX7
It was February 2004 when they told the dev to come up with a fixed function rendering and it was a rush
A long time ago
so it won't be easy
lol
in barnyard they forgot to implement the terrain in the FFP mode
the ground was just missing
halo 
Yeah. It seems DreamCatcher was not the only publisher at the time to require studios to support FF
Devs absolutely hated it
I still don't understand how the shaders work in the fixed function mode for halo lmao
like several tools pickup shaders but the game worked on fixed function hardware
You had to go out of your way to make this work, because in 2003-2004 you would have to maintain 2 completely different renderers
The game may have toggled the FF mode specifically when generic TnL hardware was detected
In Painkiller to force the fixed function path you create a shortcut with the "+hwclass tnl" launch parameter. Not a lot of people even know this
yeah I'm talking about the game's specific fixed function path
normally the game uses shaders
Well then there should be no shaders, since the very concept of a programmable shader is not known to a TnL GPU
Could the tools be picking up some junk from overlays or something?
nope
was able to toggle the skinned meshes via shaders using the renodx devkit lol
so there are definitely shaders in there
Well hardware skinning wouldn't run on a GeForce 1/2 either way so
I guess it is exposing some stuff for newer hardware
I'm having vibe take a look at it
Thats Mr. Vibe to you!!
lol
okay so vibe is saying that they are in fact software rendered vertex shaders
On a GeForce2 (hardware T&L, no vs_1_1), Halo creates the device with MIXED_VERTEXPROCESSING. Before each skinned draw call, it calls SetSoftwareVertexProcessing(TRUE). The D3D9 runtime then executes the vs_1_1 shader bytecode on the CPU โ Microsoft's own software vertex shader emulator, built into d3d9.dll since DX8. The same CreateVertexShader handle, the same SetVertexShader call, the same bytecode โ all identical to the GPU path. The runtime just routes execution to CPU instead of the GPU shader units.
Debugging and tracing tools intercept the D3D9 API layer, so they see SetVertexShader being called regardless of whether the underlying execution happens on GPU or CPU. There's no distinction at the API level โ it's invisible to tools like RenderDoc, PIX, or your own tracing hooks.
lol looks like the sims 2 legacy collection does something similar: https://discordapp.com/channels/1028444667789967381/1482811180870406376/1486964743423004762
What if
this model is a proxy or something along the lines of that, as in 1 model that houses 3 other fake ones?
cause like I noticed when I switch it into the tag terrain, only one model stays left, the other 3 models disappear
I also noticed that they're mostly props (the ones that get connected)
when I also go to edit their tags, they're all the same
if I set one to lets say texture Ignore/hide, they all disappear
Ah yeah, makes sense
For some reason I was thinking that the tools would only pickup shaders executing on GPU, even though there's no difference at the API level
also makes sense since the game was primarily developed for shaders
so software emulation might've just made the most sense with the constraints they were under
Bad camera does not cause geo streching/merging. So itโs something else. Itโs a longshot but I did have something similar in my idtech3 games, I could send you a patched remix see if the problem gets fixed, I just need to put it on github
Yeah your right, I think I know why though, but do you think yours could work? (since the game uses d3d9 shader model 3) and would probably require modification before I can test it
.
A dev also told me that yeah the game does use static mesh instancing to an extent so I'll have a tinker around with the agent about that, if all goes too unwell I'll just give your patched version a shot cause why not
Cause I've tried to disable BLAS, disable spatial reassociation (for some reason), and so on, but none of that worked so its either this or the math I have is wayy off
For me it happened with unstable meshes: one moment it would look ok, then move the camera a bit and geometry would leak into the great beyond. Then itโd go back to normal, and repeat. It was the BLAS caching.
You mean you have to modify dxvk-remix too to make your game work?
(Iโm a bit anal about these things, so grain of salt.. but try not to modify remix with AI unless you have a grasp of how3D rendering works.. fix your game in the proxy)
GPT 5.5 put the shadow terminator fix into my Manhunt runtime in one go. Pretty good. Just gotta check if it's as good as vanilla

You get 7000 credits for $40 
Does GPT 5.5 usage with Codex "cost" more than 5.4?
5.5 is 5.4 but lobotomized
Use 5.4

I find 5.5 to be an excellent model
A fine build up upon already great 5.4
5.4 was good but 5.5 is great
It did most things in one shot, even complex stuff. It's just too expensive on Github Copilot
Just wanted to say. It is good to the point that I've become confident enough to expect every prompt to be fruitful
Are you using it on Codex? Is it using more tokens or whatever than 5.4? Getting to use 5.5 freely would be really OP
I never used 5.4 in Codex so I can't tell
They're saying that 5.5 is actually more efficient and I'm kinda inclined to believe that
Doing some UE3 stuff that was never done before together with a very experienced programmer and the guy is blown away by it
You got the 100 buck plan?
Can you just go ham on it for a while or what do the limits look like
Im not a 24 hour user but I'd like to be able to do 1-2 hours of work when I want to and not be too limited
A couple of hours of sweeping through the UE3 source and implementing stuff takes like 5% of weekly usage
They provide quite a lot of usage really
5hr limits always restored before running out
That sounds good. Was that last month with the doubled limits?
It was 14 hours ago
Go to whatever is good when you resub
It's a cycle
right now we're scamming OpenAI
I just hope that going public won't end the party
๐
do you think gpt 6 will be when it can do things in one go
It is already extremely skilled at working with source and better than humans at recognizing patterns and noticing things that matter
My good friend has built a very versatile and clever architecture for uncooking UE3 maps (something that was never done before. There was only UE Viewer by Gildor and it is merely an asset viewer) and GPT has been extremely helpful at finding things my friend missed in his heuristics
Unprompted at times
I will be re-subbing to Codex come June 6th, I was leaning against it for my current project. Been stuck in deep research for a solid month fixing random issues. As of tonight 5.5 went on a spree and fixed all outstanding issues. Most stable my proxy has ever been. Made the project viable finally.
Was dealing with a shit ton of geometry hash churn from quite literally 5 different sources. Half bad proxy initial implementation and half weird game behavior. A lot to research but it fixed it all one by one
Worth mentioning that Gildor worked on UE3 and had direct access to it
wait omg the dev of umodel?
Yes, UModel = UE Viewer
Epic Games asked Gildor to rename it
heโs such a goat
Our tool is even better though
Gildor's don't even recognize some arrays

Shhhh he was a goat for making viewing fortnite models so easy
like, for example, UE-Explorer cannot detect arrays which are not already pre-defined (usually game specific arrays)
I like comments in UE3 code
Epic Games employees becoming self aware that their code is shit
I also know that Ron guy bruh

Why you UE3ing this much is there a Painkiller on UE3
Yes
Knew it
There's also one on UE5
But this one is imposter, the only thing in common is the name
Wish I had a single game I cared about that much to focus on. Instead I could probably spend the rest of my life working on random games so I have to pump the brakes a bit
And it's actually on UE4, so I couldn't even tell what engine it uses
The dark side of sw dev: good developers writing shit code, and crunch being integral to the job. Also probably why so many managers are jumping on the hype train with coding llms. I just hope that weโre getting the AlphaGo progression on these coding-AIs so I can take a well deserved vacation ๐
I mean I personally know the guys behind the game. It was made like 200km away from my hometown
It's a project that has cultural/historical value to me
I don't know what should I feel, happy that ChatGPT nailed it and did it better, or the fact it reinforces the idea that it is better than employees

I remember how proud I was when the ubi studio in my country got listed in AC Black Flag credits 
(Alot of ubi studios worked on AC BF ๐)
Gildor worked at Ubisoft on The Crew and said it was a good company to work in
It was in 2013 though
The fact that this is the game that kick started the Stop Killing Games initiative is ironic
Iโm actually hopeful, in the 50s brainy techbros were writing assembly code. Then high level compilers came and prolly no one trusted them. Today we are almost certain asm code from a compiler is better than anyone could write.
We would shift the paradigm, ie write specifications, and it will be more accessible, for design oriented individuals, but there will still be employees.. but no longer working in the โminesโ of sw development
I'm also very much excited, but I'm not the one who needs to answer "What are you better at than AI" at the interview
Because I'm not writing code for living
We have interesting times ahead of us that's for sure
That's how it always is
Poland was in such a doomed state in the 1990s that Belarus and Ukraine were better places to live
Poles went to Belarus to buy food
Now look where it is now
It's not exactly the same, but it reinforces the fact that decisive economy changes are for the better in the long run
And that's it
Only lasted a day 
Throws you out mid response too
MS can't afford to even let the AI finish a prompt
Wallets be empty
Fun fact
Codex finishes processing the prompt first before running out of usage, so it's not leaving you with a broken code halfway through the work
Yes, and I had an impression that this is kinda a standart
When Opus just said "see you space cowboy" and left me with broken shader - it was anticlimatic
I mean prior to the usage based billing Copilot also worked that way
Just now they pretend to have to be hardlining it
You got people at Github crying about how unsustainable it was costwise as if they weren't the ones to put it up at that price. And then they went from it being a steal to you essentially paying directly for HW in a data center somewhere
Fun fact
We needed to bring some shader from NecroVisioN (DX10) to Painkiller (DX9) so we used GPT 5.5 to backport the Shader Model 4 pixel shader down to SM3 and it worked flawlessly
The funnier fact is that it looked better than the developers' original DX9 version
Like going from $40 to thousands if you kept using it like before
The original NecroVisioN DX9 shader the devs made themselves
Looks like crap
That is... Quite impressive!
Dam, I haven't played Painkiller for like 20 years...
Need to fix that one of these days
The fact that you haven't played Painkiller RTX is absolutely mind.blowing
I mean we won $40000 brother
I have somewhat strick policy about unfinished stuff
Games, mods, TV series, book series, ect
And from what I understand Painkiller is still being tinkered with?
So it is VERY MUCH in my logbook
Always
Still it is very complete
Hmmmm
Hmmmm
I do need some nice entertainment before Victoria 3 gets fixed...
I won't say that it is Portal with RTX level
Eh, don't think anything is
It was too deliberately professional to be considered a yardstick
Is there anything like a "end vision" for it btw?
Or as long as Remix gets new features, Painkiller will be poked with?
We've built a solid ground for Remix in this game and have some absolutely crazy programmers at our side
So I would dare to say that it is one of the best games to test new features on
Oh, I absolutely don't doubt the quality)
what optimization techniques have u implemented
I am just a sucker for "this is finished" labels, for better or worse
But yeah, def gonna try soon if it's fully playable trough
We switched from AABB culling to expanding the nearclip of the frustum and moving it behind the camera. AABB was already great and this approach is even more performant
Also we're not rendering static geometry anymore, it's only in Remix now
whats the perf like
So that the game only renders dynamic meshes
Huuuh, so half of geometry is game-side as usual, but other half is basically Remix-native?
Much much better than it was before, like A LOT. Talking 3x improvement here
120 fps 4k native obv
The problem with Remix is that it has some wild CPU overhead. Eliminating static geometry makes it so much better
As most of the geometry is static, almost everything is only rendered by Remix
imagine the perf if the static BVH was pre calculated along with sharc
@solemn epoch โค๏ธ
It's always nice to see the research press on and new features get implemented
Really looking forward to what the future tech will bring
ya this is all in ds2LE rn
where at the best of times fps can reach 80-90 fps on a 5080 1440p
so cool
Yeah I dont think thats my issue tbh cause I once disabled BLAS and it still didnt work, as for the modification, aren't you supposed to modify the dxvk if your game anyways if your game isnt supported etc? cause I never heard anything about a proxy
The FFP proxy is part of the toolset and it sits between the game and Remix to do all the conversion you need to make the game compatible
Only modify the RTX Remix runtime if you really, really have to
For compatibility I mean, if you want to add features feel free
Installed and all but Im running into this issue when trying to build the repo
D/R, he fixed it
๐
Today Microsoft moved all its GitHub copilot subscribers to token-based billing.
If you want to see what happens when people have to pay the actual costs of AI, the day is finally here. It's obvious that every customer sees the deep, meaningful value and isn't angry at all.
3292
9870
865
319
yeah I hit 50% yesterday 
time to wrap up anything i have left for BetaRT and just ship it as incomplete ๐ซ
how nice, came right around when I was using copilot, now I'm locked out for a month lol
windsurf is agood one
you'd have to improvise though on the instruction for the VRET
havent done that yet
yea im still paying for it as a backup
it wont be long until it follows github copilot
while trying to use it on copilot i got an error about d3dx9 or something
prob the d3d9 sdk?
LINK : fatal error LNK1181: cannot open input file 'd3dx9.lib'
ye
yeah im not sure on what to do tbh
I installed the 2010 runtime but that didnt work
install the directx sdk and make sure its part of ur path
runtime isnt enough
This download is the legacy DirectX SDK. It is now end-of-life and should only be used for maintaining older software. For targeting DirectX for Windows 7 or later, use the modern Windows SDK instead.
thanks
People got to stop going to these 3rd parties, just buy direct and buy the $100 plan. If you dont have that ability, then prepare for the $20 plan to be a demo
they're really not that much cheaper, and this token pricing model is 99% likely to go to the first-parties
the bubble is about to pop
hopefully more pressure for local llms to become better (obviously this will take maturity on the hardware-side and software)
Nvidia Vera Rubin architecture entered full production. Deployment later this year. It is 2.5x faster in FP4 (AI calculations) than current Blackwell
Can't wait to see GPT 5.5 but 2,5x powerful a year from now
i dont think that's going to be relevant for any consumer if hardware costs dont go down in the next 5 years
my main worry is that it wont be a quick pop but a painfully slow deflating
What do you mean? Every time new Nvidia arch releases there's a huge jump in AI around the corner
ok but who (cares) and who can even take advantage of it other than the 0.01% of the rich?
but no one can afford it is what im pointing out lol
the only realistic way for local llms to prosper is getting better performance/content out of hardware the majority of consumers already own
I mean they are getting better at that too. There's a heavy 1-bit models research going on. But the real push forward is always brought by the new generation
what do yall think of the spark announcement? https://youtube.com/shorts/coKZy1VMUc8
What is Jensen even talking about anymore?
#computex #computex2026 #nvidia #gaming #n1x #geforce #ai #bubble
Patreon: https://www.patreon.com/MooresLawIsDead
US Amazon Affiliate Link: https://amzn.to/3ABOcPl
Cash App Donations: $MooresLawisDead
Twitter: https://twitter.com/mooreslawisdead
sure, maybe i didnt explicity word it well, but yes newer hardware would be technically the faster way to get better local llm experiences, but i just dont see anyone wanting to buy this stuff with how expensive it is
Likely the price of a 5090 with the performance of a 5070
Nah its rumored to be between 5060 and 5060 ti perf
so even worse value
"No no no you are wrong its worse!"
its literally almost near the perf of the existing DGX spark
It is just the dgx spark chip
but placed in a laptop
People always buy. The tech will never cease improving and there will never be lack of interest in it
the people who can afford it*
That's not really fair
It is expensive, sure, but it's not 10x expensive
It's not even 2x expensive
You can argue like this about everything in the world
Doesn't make much sense
the main worry is a terrible future where itโll reach a point where the 99% stops buying in focus for neccecities
People experience economical challenges for once in their lives and think it is over forever and ever
PC hardware has it good I can tell you that
That we can't buy materials for the fucking house that's a problem

also that dude just going โ15-20 years ago we had an idea called a phoneโ
โtoday we have an idea called a pcโ
Full stop
what does that mean
are we sure about that? I dont have the exact numbers but this is the price of the ram i pad for in 2024 vs what it costs now, it doesnt print a pretty picture
it is asically 1 to 1 desktop 5070
RAM - sure
But that one is overpriced even today, can realistically find 64gb for something like 700$
And RAM is only part of the price
any type of ram, VRAM, dram module themselves, you name it, its gotten at least 2x expensive
everything tech relies on that
im guessing its perf for the power limit its set at
i don't think its nessessarily fair to act like that's a temporary problem until we actually see price drops with the decline of frontier AI usage and rising subscription costs of those services
around the same for mine
my SSDs too
the SSDs are 2.5x what they were
RAM is like 3.5x
yeah i can grab the receipts for any of the SSDs i bought as well
๐ same ssd
i got mine in feb 2025
I should have got that fuckin WD Black 8TB
I can definitely see where all the memory goes. AI has experienced an absolutely wild breakthrough over the past year, I think no one is going to argue with that. OpenAI only expects to become profitable around 2029, so you can expect the heavy investment in the AI research to press on until then
I see this as a shock therapy of sorts
We need to power through this
The AI research is bound to slow down and the demand will eventually plateau
We're literally almost there with all the agentic progress
I don't see them becoming profitable given their current and past business decisions, just my personal opinion though
only the USA government can help them lmao
with that pentagon money ๐ค
or palantir or whatever
you see, all the money goes into research. They've been doing it since 2015. Once the foundation is there and integrated into most parts of our lives, there's no turning back
It'll become the norm, the new economy
i dont care about turning back, i care about my shit being affordable
and local llms thriving so i can stop giving altman my money
You must give it some time
Still hoping for China to drop a model that beats they ass
We're in the transitional period
yes 
i pray for chinese saviors every day 
Running on Chinese GPUs ๐ ๐จ๐ณ
Once the Chinese government makes AI infrastructure something they want to have it'd go fairly quickly
my computer WILL become the property of CCP and that is a price I am willing to pay ๐จ๐ณ๐จ๐ณ๐จ๐ณ๐จ๐ณ๐จ๐ณ๐จ๐ณ๐จ๐ณ
The point is, inference itself is not that expensive. Research money shilling is
They constantly build new datacenters because it advances so fast
LOL
cannot wait to arm the homeless to raid these datacenters


I wish I had a explode all datacenters button
me too
Then we would have to start over and suffer another couple of years with these prices
if anything a spawn new datacenters button would be actually helpful lol
I'd just keep pressing the button 
it exists but I cant tell you about it

Press it mayne
Probably signed NDA for that

Who had that one book that seemed useful for ffp
virgiliojr94/book-to-skill: Turn any technical book PDF into a Claude Code skill โ ready to study, reference, and use while you work. https://share.google/1xkuK7QOlTOwEYM5t
???? Found the smoke test?????
@sharp seal ? Was it you ^
๐ฌ ๐งช
A YouTuber with 110 million subscribers released a free version of ChatGPT.
His name is Felix Kjellberg. You know him as PewDiePie.
He spent his own money on a 10-GPU computer at home. He used it to run the same kind of AI models that power ChatGPT, but on his own hardware. Then he wrote his own app to chat with them, because the apps that already exist were not good enough.
Then he gave it away for free. Anyone can download it. Anyone can change it. Anyone can run it.
It's called Odysseus.
It runs on your computer. Your data stays on your disk. No account. No tracking. No monthly fee.
What you get:
- A chat window like ChatGPT
- An AI assistant that can browse the web, read your files, and do tasks for you
- A tool that scans your computer and tells you which AI models will work on it
- A research mode that reads many websites and writes you a report
- A side-by-side mode to test two AI models on the same question
- A writing editor where AI helps you,โฆ
Normies discovering local models lmaooo
Lolll
well
The twitter post is so dumb
im pretty sure most local models take up so much RAM/VRAM no?
Yes
and to output stupid stuff
fair enough
this isnt a result of trying to get it to say "smoking gun" on purpose (Like SOME of you do who will not be named but it is obvious) it just started saying smoke test on its own????
Smoke test is an actual term
Nothing to do with smoking gun

Well it was funnier when I thought it was making it up lol
I guess that makes sense
?
i mean hey at least pewdiepie is using his influence to push for local models to the masses so we get to a future where local ai is small and usable
even if rn its a bit dumb
all takes time
linkedin coded ya
LOL
well
I only changed the resolution and the game locked but like worm???
my anti virus is going crazy on innocent files
Itโs been an issue since forever. Because remix, in order to hook the games (capture mouse and kb input), does things similar to what a virus/malware would do
oh..
im convinced "compact conversation" is completely fake and made up, it just starts a new session and sends the previous message you sent. It's just so bad
Ooooh shots-fired: Today, I AM THE ONE who sees the smoking gun! ๐
Bro really complained about how they treat him
To get the best results next time, I spent time analyzing your message for inefficient usage.
To be fair, I usually go to Gemini pro and say "Hey you know idtech 3 right? Here, I modified this function and now is fucked, plz fix". Gemini is a boss and does it.
Now I'm rehashing the same convo with local models, and when the models fumble I go "Admit it, you know nothing of idtech 3! How do we work together better so that you do more work? Help me help you!" but politely of course, I am not a peasant.. and well.. I got my just-deserts today..
But the response was fair: Since it does not have enough knowledge, it's asking to create a local database for the codebase, then I should ask it to fix the booboo I did
It was super fucked when I tried DFX2, and claude claimed the whole world was made of skinned meshes
Or CPU skinned or some whacko shit that made me drop it pretty quick
Also still suffers from FFP issues
Claude also claimed this of Deadly Premonition but was wrong
No one likes dfx2 and you're wrong for liking dfx2
It is objectively, in all senses, a bad game.
Delta Force Xtreme 2 was revived by the community after it was shutdown 2 years ago!
To get the multiplayer fix, head over to the BCU website, there is on the home page an image of DFX2, just click on it and it will take you to the download.
www.bcuveterans.co.uk
โฆ
esmoke
If Anthropic bans Smoking Gun in the system prompt It'll qualify as Opus 5
Neon and brake lights
Holy fuck I think I might have found a way for OP light placement
a draw with texture T and vertex-colour C โ emit light L
Use the game's own light flare system to place acurate lights in the whole scene.
It's kinda similar to remix's "Add light to texture" system but proxy side for better categorization, customization and stability.
Interesting. I guess your problem would be: How do you decide where to position said light source? Center of the mesh aabb? object pos? Some are positioned at 0,0,0 so you need multiple ways to customize such authored lights 
Center of the sprite aabb works for me
All the flares are exactly at the light's source
If a light needs to be rotated, I can set a world axis rotation constant
Thjat's for later though
I'll test with the nitrous particles firslt
Gotta be cautious with the same texture being reused in many places (I.e: rain dropplets)
They are separated by vertex colors
So all the light flares are b/w
I find the drawcall. Separate by color.
Got it. It is quite niched but sounds like a good plan for U2 
I only tried it because it was the latest game lol
I dont think I even tried DFX1
That's something I would write to the AI
I never saw it misspell something but it did abbreviate We Love Katamari Royal Reverie to "KatamarirRR" with an extra r which was really embarassing 
Me misspelling embarrassing in a post about misspellings is good
"ive hit a frustrating snag" SHUT YOURSELF DOWN NOW







