I really enjoyed that salad! Want to start making money in your PC's downtime too? Get the Salad app here bit.ly/Kaizen-Salad and start Chopping today! Make sure to use my code UG5JA0 for a 2x earning rate bonus.
Just one detail about decreasing the tile size: It helps you to render big scenes when you don't have much VRAM available. With a powerful GPU it's generally unnecessary make it smaller but when you can't render in 2048 tile size, decreasing it would allow you to render in exchange of more time actually. This is my experience messing around with it xD
3080ti: i always deactivate tilerendering. I often render in 4k and it's always faster than keeping it on 2048. Only when i push the gc to it's absolute limits i turn it back on.
+1, when i was rendering on 1660ti i constantly used tiles with numbers like 768-920 to speed up render. With 4090 using tiles is only increase render time
Good to see Noise Threshold being properly explained. So many people still just parrot "lower your samples" when this has not been the primary control for render quality since 3.0. They even renamed it from Samples to Max Samples to show it's defintiely a different setting, but no the parrots just started saying "lower max samples". Lower max samples does one thing - reduce the samples available to the noisiest part of your image. Exactly what you don't want to be doing.
Which could be the standar setup for using NT, because I used it in 0.5 with 3000 samples and the results looks way noisier that a NT at 0.01 with 300 samples (And for animation means more flickering)
Some tips for faster renders. This is coming from someone who renders stuff at work on an ancient PC with no GPU only an or 9th gen i7 CPU. 1. High resolution (~3000x3000) + low sample size (~32) + Noise Threshold 1. This is really good for low end PCs. 2. Small tile sizes (~128 for me) are good for CPU rendering, while larger ones a better for GPU rendering. 3. Sometimes adding textures after in Photoshop is actually faster. If you have a large wall with a lot of bump-mapping.... just do it in post. I can pump out simple scenes under 5 mins, more complex stuff under 10-15 mins.
A note about adaptive sampling: It's worth trying out a scene with a lower sample count and lower adaptive sampling value in complex scenes, as it seems that the AI/algorithms were set up for images with uniform sample counts. In my experience the varying sample counts from this can cause artifacts in the denoiser. To prevent issues try increasing the min samples near to a quarter of the total samples and messing around with values as always. I usually use a value between 0.01 and 0.05. The second image at 14:50 looks like it uses a smaller sample count or some other setting on top of disabling caustics, but it's also possible that the reduced noise from the lack of caustics is causing the adaptive sampling to give it less samples and therefore increase the artifacts in those areas. On top of disabling the animate seed option that someone mentioned to reduce flickering, using the automatic scrambling distance option can also help reduce noise in renders by allowing nearby pixels to have more similar random values. Setting the multiplier to 0 is also a great way to visualize converging times and figure out what is a good sample count for you! Transmission bounces are required for any transmissive materials, that includes glass that uses transmission, refraction, etc. Transparency bounces are only used for materials with alpha (i.e simulate a transparent material with no internal bouncing and only a surface) or using a transparent BDSF.
BVH trees have to be recalculated if the triangle world-space positions change so there's only so much data that can be persisted frame-to-frame. Edit: The wikipedia article on the subject offers a decent albeit simplified explanation of the subject.
As a longtime blender renderer and dabbler, this video was FANTASTIC in being visually captivating, well narrated, and hey, still learned a few new things, too! Subbed.
You dont know how much I appreciate you making me find out my render device setting was on NONE all this time, I knew it and I never noticed it, I feel so stupid but I'm glad I watched this today cause I have to render a lot of work tomorrow
You missed something while talking about denoising. Under advanced settings in Cycles sampling panel there's animate seed option. That one needs to be disabled in order to avoid flickering between the frames. In fact, quite important. Also, I haven't seen anyone mention this but since 3.6 version, maybe even one or two versions earlier not sure, there's Turbotools add-on inside the Blender. That's accualy separate denoiser. If used properly can give even better results that Optix or OpenImageDenoiser.
That's not free though. I still use the ai denoiser from recycled. I've separated the addon from the special build and still jas way better results vs blender vanilla approach
Even better; apparently there is a temporal denoiser using Optix in Blender. But it's hidden somwhere under the developer extras. But yeah you're right on the seed for the noise threshold, good one!
@@KaizenTutorials the only place where I saw Optix temporal denoiser is inside the free Pidgeon Tools add-on. I've been playing with it for a while. Interesting results. Check that one. Cheers!
Dude I've been rendering these animations for so long taking hours of my day just for a 5 second animation just to figure out I've been limiting my hardware and could've saved hundreds of hours just by clicking a button, my life is a lie
About using your CPU in addition to the GPU for rendering: It depends. Usually, it *should* be faster, but there also are a lot of cases/scenes where it actually slows down the rendering, in comparison to a GPU only render. Always test which is faster before rendering your animation
On laptops, CPU+GPU is the worse option almost 100% of the time... Because either GPU or CPU alone can max out thermal/power constraints, it all boils down to how many samples per watt the system is producing.. By taking this some of this power budget away from more efficient GPU and giving it to far less efficient CPU will result in less performance.
A few things to note: Make sure you are not thermal throttling when rendering with both the cpu and gpu (do not do this if you have a laptop it will reduce render times). Noise threshold can cause artifacts with denoising (especially with volumes). Experiment with the minimum samples to make sure you dont get these. Often times i find it best to just disable it entirely. Persistent data does not work with adaptive subdivision. I also find that keeping render times the same, rendering at a higher resolution results in more detail since the denoiser has more pixels to work with.
Thank you so much for these tips. I've been trying to render out a 1053 frame animation, but each frame was taking over a minute to render, even with the noise threshold set to 1 and samples reduced to 1024. I found the Optix option, played with the lighting settings and a few other things, and now, from my current test, each frame takes around 9-11 SECONDS depending on the complexity of the frame. Massive time saver ! ^_^
This Video helped me a lot for getting happier in Cycles. At least for me denoising settings were somewhat of a Problem. After turning off the denoise for the viewport i got much faster results in simple scenes. The best results in terms of speed were achieved by turning Denoising off, but if you have the extra 2-5 Seconds in the viewport to denoise i recommend to set start sample to the last frame. This way not every sample gets an own denoising process, but only the last one does. So you wait less inbetween the actual Samples. Haven't tried this in more complex scenes, but for hardsurface modeling this works like a charm.
Dont know if anyone said it before, but if you use a bigger format, like 4k instead of 1080 and use 1/4 of the samples the denoiser is able to mantain a lot more details and sharpness. You divide the samples by 4 because 4k is 4 times the area of 1080p. Max Hay uses this technique and downscales the render to 1080p afterwards. You don't actally have to change the resolution since changing the format % to 200% should be the same thing. Hope it helps someone! P.s. Please do a full video about temporal denoising techniques and fireflies/artifact removal :P Loved the video from Southernshot, but I was asking myself if there are any other workarounds (maybe even in AfterEffects or premiere)
it could be. It could probably be 1 min long, but there wouldn't be any fun or explanations in it. And I'm a strong believer of not just showing, but also telling!
Another tip: render from the command line. Some utilities online can do this, for those unfamiliar with shell scripting (like the Flamenco add-on, which I use), but this is another way of managing memory. This way, you don't have the full Blender UI running in the background while rendering - helpful for longer animations. A middle ground to this is to lock the Blender interface while rendering (meaning you can't use the UI while it's rendering) by going to the top menu -> Render -> Lock Interface.
Yeah i've heard it can shave off a couple of seconds, but I honestly couldn't get it to work myself. Then again I have 0 coding experience and understand nothing about it haha
Here's my rig, it's an gaming rig tho my workstation is using Ryzen R9 5950 + RTX 3090, im sure you already know the performance in Nvidia side so here is the intel ARC GPU for you instead. Intel i5-10400 + ARC A770 16GB Render time : GPU + CPU = 8m02s GPU Only = 6m29s note : since it's taking too long to render at 4000samples I reduced to 2048 samples and noise threshold to 0.03 So clearly if you're CPU is slow it will actually hinder the performance.
Ah yeah so it seems, thanks a ton for sharing these numbers! As for the intel ARC gpu's; i'd hoped for them to be faster. Using NT 0.03 and 2048 samples takes my machine around 1 minute.
Yesterday I actually found out about these and my rendering time went from like 15 seconds to literally like 2 or 3 seconds (it was not a huge render thing but just a small one but doing this stuff helped 100%!)
One thing I noticed is that some of my colleagues or students are using downloaded materials (from Blender Kit for instance) that are way too complex for no reason, like a glass with 3 different IORs and refraction values, or a shader with multiple 4D noise textures with the detail slider set to 15. Also, one of my student had 8 hours render time, I was able to cut down the render time to 10 minutes just by some optimization, for instance some tiny objects had a PBR setup with 4096*4096 textures, completely useless ! In addition to what you said in the video, these are some important things to check : - Polycount - Texture sizes (also if a texture is 5000*5000px for instance, resize it to 4096*4096) - Heavy procedural textures (the "detail" slider doesn't have to be set to 15 all the time) - Volumetrics are killing your render times on Cycles (something like 10 times longer is normal) Also, I noticed that Blender doesn't like to have too many objects (i'm talking about thousands of objects), if you can join them it can make a difference (or use instances if they are similar) Thanks for the video !
Biggest mistake I recently made was a total rookie move. Rendering a simple animation in cycles, but each frame was taking almost 2 minutes. Finally realized I just forgot to change the device from the default CPU to GPU. Went from about 2 minutes per frame to under 20 seconds! Sometimes the simplest setting makes the biggest difference.
Not me waiting for my 350 frame animation to finish rendering while watching this video 😅. Usually, I significantly decrease the sampling and resolution, so these tips are a huge help
If youre using GPU, deactivate Tiling, Turn on Denoiser Optix or OpenImageDenoiser, low the samples to 12 and in output in resolution activate 200% With this you will denoise a bigger image that will have more details and you will get a better result with the denoiser because this one understand better the image. And deactivating the tiling, its because GPU handle better the whole image than tiling images. Doing test the tiling always got more time than the whole image.
I’ve seen a few people recommend that low samples high resolution trick and I’m a bit skeptical about it. At least in the scene I’m testing that has volumetrics, the results are quite a bit worse and has very noticeable blotchy denoising artefacts in dark areas. Not to mention the slightly slower render times even at such low sample counts and you’d be using exponentially more storage space if you keep the renders at that 200% scale. Doesn’t seem worth it to me.
To avoid running out of memory, I always save the project, restart the pc then render immediately after its on. because restarting your pc will clear all ram and that's what we want.
AFAIK, Tile size optimizations only works if you have more GPU cores than what cycles can use with parallel processing in a single tile. If you have equal to or less, then you are best off rendering 1 large tile. If you have more, your better off rendering a number of tiles that has the largest number of cores per tile simultaneously without going over. at least, that's how I understood it. I don't know how many cores cycles is optimized for. if it's 4 cores, and you're doing CPU rendering with a threadripper with like 60+ cores, multiple small tiles will be faster because it's rendering multiple tiles at a time, but AFAIK the average consumer-grade GPU can only render a single tile at a time, and reducing tile size and increasing the number of tiles, actually increases render time, rather than reduce it, because it's only 1 tile at a time, and switching tiles takes time to load new data to the GPU. I am NOT a programmer, or really any kind of expert, I just watch a lot of youtube videos, and tried to render on my laptop and this is what my experience of it points to. simultanious tiles = total # of cores MOD max cores per render.
So from what I've been seeingin the comments tiling is mostly useful to prevent VRAM issues and for CPU renders. Modern GPU's render faster on full tiles vs. small tilesizes.
There is mistake in the description of how Light Paths works. Transparent is not responsible for bounces in the glass, it works in bounces of the alpha texture in the shader (ex. leaves in trees with alpha mask)
You're right! But that also affects glass believe; atleast it's transparent component. which you can see in the render as using these bounces did make it see-through again. But for the light refraction inside of glass you'll need other bounces!
@@KaizenTutorials You can doublecheck it with array of glass planes on front of Camera. If you add alpha weight to you glass shader, transparent bounces will work
I was told several times earlier to deactivate CPU from Optix, because it "gets in the way", though both my processor and graphic card are not as good as yours. Also the scene is amazing!
Ah yeah it seems to depend on the hardware honestly. On my current and previous setup (3070 and ryzen 3700X) using both was faster though! And thanks, appreciate it!
Hey ! Thanx ! Love your videos, they are helping me so much and I also like the style. Don't mind what some individuals with ridiculously empty calendars say in the comments.
I always crank the render settings to 1024 tile size and 2048 samples to then denoise outside outside Blender or Maya. I bought a 3090 four days ago and after coming from a 1060, 2060, 4060... it's a breeze. Something funny is that the 3090 is much more accurate at tracing than all of the GPU's i used so denoising is very easy even with few samples.
honestly before watching this video I was thinking of just sending my projects to render farms becuase my blender would crash immediately, i am used to just modeling and rendering a single frame for small simple assets so I dont save my project till the very end, my mistake became obvious very quickly when i noticed my whole project was gone, including all my assets I made in the same project, dont do what i do...
For tile size: It did massively help for me on the CPU (especially in multi threaded systems), but in my experience, it's comparatively slower on the GPU
I love those fast renders. At one point I didn’t know exactly what I was doing and I would render a billion frames at an overly large size. It would take an hour to render when really I needed one square frame. 😅
Step 3 is completely wrong Optix is fast af, but Openimage is huge WAAAAY more better quality. And to get rid of "Denoising artifacts" on animation you just need to set noise seed (under advanced tab) any number than 0. So the noise pattern would be the same every frame and there will be no denoise artifacts on animation and the animation would be smooth ASF
In my tests Optix performed about 90% the same as OID for stills, with OID being the clear winner, for sure. But for animation I don't like it at all. The seed is a great tip, but apparently there's even a full blown temporal denoiser available in the developer extra's right in Blender!
Honestly I barely use the noise threshold because I like the more obvious result of just increasing the number until it's too slow. I feel like the not noisy areas don't take long to render anyway
I don't wanna be a shill, but there is a addon I use to use a hell of a lot called "Super Image Denoiser" which does have a temporal denoising setting. Back when I used it it could do a pretty amazing job at denoising scenes, but nowadays I'm fairly sure it's broken as it's free version (I think) has stopped development and is now been put into a pretty nifty addon tool kit
another way to speed up renders is like they use in vidddygame engines is to compile all them texture node to trim sheets and hero textures , that way the compile time is in milliseconds , also like show'd with single color backgrounds is to use the just a diffuse node not a whole procedural node when needed , also a make pass renders for each of the diffuse , glossy and transmission and then can use a compositor later to fix the colors and intensity as well as speeding that by dividing scene into collections for foreground , mids , background and holdouts
Thanks for this! I had everything set like you did minus making both CPU and GPU run at the same time. Changed that one setting and my 240 frame animation went from a 45 second render per frame, to a 5 second render per frame
I was alway using OID for animation because I heard in the past that it was better for animation. I Guess it's not the case anymore, gonna try optix Denoise now.
Quick heads up about 2:53 : Render times can still be slower if you have both CPU and GPU enabled. I tested it in a scene I was working on and it was quite a bit slower (I'm using OptiX).
Got the rendering time to 59,75sec with rx 6600, ryzen 5 3600 and 16gt. Gpu only, no tiling, no noise threshold sample 64. OpenImagineDenoise accurate, seed 1, max bounses 5,4,4,4,0,8, Fast GI, Simplify 1 max divisions 12 & texture limit 8k. Use spatials, persistent data.
Dude is this true ? “KAIZEN™ means improvement. Moreover, it means continuing improvement in personal life, home life, social life, and working life. When applied to the workplace KAIZEN™ means continuing improvement involving everyone - managers and workers alike.”
Enabling OptiX only for my GPU, and not for both CPU+GPU, improved my render times, from 12min to 9min And my CPU is not at max usage now (7800X3D and RTX 4080) Edit: Well Blender disable your CPU for OptiX if you have a gpu enabled, at next startup
to tile sizes: by setting your tile size to the amount of bits in your GPU it can shave off multiple seconds (if you don't know your bit-count, use 256) i don't know much about this but it pretty much just renders the single squares and if the GPU doesn't need to cache anything it'll be faster i guess
u know i never played with the noise threshold before this . . . u have saved so much time on renders . . . from 5 min to under 1 . . . with openimage its still fairly good unless zoomed in on reflections
faster isn't necessarily better however and you can then find yourself fighting artifacts you don't understand as a beginner because you followed a make blender render faster tutorial. I just wish these videos had a bit of disclaimer in that regard.
In my experience, tile size just puts less strain on the hardware. I kept the tile size the same while I was learning blender, but kept getting crashed when trying to render multiple donuts. Now Don crash yet
Rendertime with Blender 4.0.2 for the original scene from the link in the video-description, frame 1, no changes made: 05:25:27 minutes!!! Hardware: AMD 3970X Threadripper (32 Cores); ASUS ROG-STRIX RTX-4090 OC; 128 GB RAM; 2 TB Samsung SSD 970 Pro; OptiX, Cycles CPU+GPU-rendering; But turning down the samples and the threshold will make it faster for sure. Rendertime with changed parameters: Noise Threshold 0.05; Bounces: 8, 4, 4 ,8 ,0 , 8; Persistent Data; Max Samples 2048 equals 00:31:21 minutes!!!
you should turn off the CPU and only use the GPU, and the default sample rate of 4096 is unnecessary high if you try to render a single image. What do you mean with original scene?
@@MrDragon005That's what I never got a clear answer for at discord-channels: Is it better to use CPU AND GPU, if having a "not sooo slow one", like me, or use GPU-only? Please explain your opinion. :) "Original Scene" means, that I have not changed anything, in the scene, before hitting "F-12". :D Rendertime is 00:28:14 minutes if using GPU-only!!! WHY????
Using 4096 is fine because changing the noise thresholds prevents unnecessary sample renders. So you can leave it up. As I show even setting the samples to 20K doesn’t matter because the noise threshold never let’s you get there!
ANOTHER TIPS: u can use ai upscaler (CupCat AI upscaler) and render with half the resolution desired. u can use ai frame interpolation (FlowFrames) and render with half the number of frames
Yep, good tips! They tend to work a lot better for projects with slow, predictable motion, but if you have one of those it will save you hours and hours!
I really enjoyed that salad! Want to start making money in your PC's downtime too? Get the Salad app here bit.ly/Kaizen-Salad and start Chopping today! Make sure to use my code UG5JA0 for a 2x earning rate bonus.
That salad thing sounds pretty cool... It's a shame that i have an RTX 3050ti... :D
You can always try! Not sure if it doesn’t work 🙏🏻
We xan actually do temporal denoising since 3.1 but is not exposed in interface
Antivirus sees this as a threat.
that's weird! You need to give it approval for firewall passage, but that's because it uses internet, but that's it! @@andriespieterse1584
the best rendering setting is an expensive GPU
It does go a long way lol
Uhm, that setting is greyed out in my bank balance...
@@GinnyGliderhmm.. maybe a bug did you report to blender ? Or try closing and opening blender again ..
Fr
@@GinnyGlider Yep, funny thing i tried this setting but received a defective one in the mail 😂🤣
THANK YOU! It really works, went from 7 minute renders to 8 second renders. Every time I have long renders, I come back to Kaizen
That's so nice, thanks for sharing!
Just one detail about decreasing the tile size: It helps you to render big scenes when you don't have much VRAM available. With a powerful GPU it's generally unnecessary make it smaller but when you can't render in 2048 tile size, decreasing it would allow you to render in exchange of more time actually. This is my experience messing around with it xD
That’s a great tip! Thanks for sharing this for the low VRAM gang!
3080ti: i always deactivate tilerendering. I often render in 4k and it's always faster than keeping it on 2048. Only when i push the gc to it's absolute limits i turn it back on.
+1, when i was rendering on 1660ti i constantly used tiles with numbers like 768-920 to speed up render. With 4090 using tiles is only increase render time
thank you! ive run out of vram very often didnt know this fix existed
3060 user here, thats the least of my problems
Good to see Noise Threshold being properly explained. So many people still just parrot "lower your samples" when this has not been the primary control for render quality since 3.0. They even renamed it from Samples to Max Samples to show it's defintiely a different setting, but no the parrots just started saying "lower max samples". Lower max samples does one thing - reduce the samples available to the noisiest part of your image. Exactly what you don't want to be doing.
Haha yes! Finally someone who truly understands! Well said 💪💪
Which could be the standar setup for using NT, because I used it in 0.5 with 3000 samples and the results looks way noisier that a NT at 0.01 with 300 samples
(And for animation means more flickering)
Some tips for faster renders. This is coming from someone who renders stuff at work on an ancient PC with no GPU only an or 9th gen i7 CPU.
1. High resolution (~3000x3000) + low sample size (~32) + Noise Threshold 1. This is really good for low end PCs.
2. Small tile sizes (~128 for me) are good for CPU rendering, while larger ones a better for GPU rendering.
3. Sometimes adding textures after in Photoshop is actually faster. If you have a large wall with a lot of bump-mapping.... just do it in post.
I can pump out simple scenes under 5 mins, more complex stuff under 10-15 mins.
Thank you for sharing your tips and insights for people with lower-end PC's!
tile sizes no longer matter in cycles X iirc. It's mostly used to save RAM now
@@tabiasgeehuman Doesn't matter if you are using a GPU, CPU or both?
@@Ragmon1 yep afaik. I don't notice a difference on my rig at least (GTX 1660 Super and i7-10700K)
@@tabiasgeehuman Well I'm gonna need to try rendering on monday, without the tiles active. I'm very curious now.
A note about adaptive sampling:
It's worth trying out a scene with a lower sample count and lower adaptive sampling value in complex scenes, as it seems that the AI/algorithms were set up for images with uniform sample counts. In my experience the varying sample counts from this can cause artifacts in the denoiser. To prevent issues try increasing the min samples near to a quarter of the total samples and messing around with values as always. I usually use a value between 0.01 and 0.05.
The second image at 14:50 looks like it uses a smaller sample count or some other setting on top of disabling caustics, but it's also possible that the reduced noise from the lack of caustics is causing the adaptive sampling to give it less samples and therefore increase the artifacts in those areas.
On top of disabling the animate seed option that someone mentioned to reduce flickering, using the automatic scrambling distance option can also help reduce noise in renders by allowing nearby pixels to have more similar random values. Setting the multiplier to 0 is also a great way to visualize converging times and figure out what is a good sample count for you!
Transmission bounces are required for any transmissive materials, that includes glass that uses transmission, refraction, etc. Transparency bounces are only used for materials with alpha (i.e simulate a transparent material with no internal bouncing and only a surface) or using a transparent BDSF.
Well said, thanks for this in-depth comment!
The BVH is the bounding volume hierarchy. It allows triangle searches in path tracing to go down from linear to logarithmic time.
BVH trees have to be recalculated if the triangle world-space positions change so there's only so much data that can be persisted frame-to-frame. Edit: The wikipedia article on the subject offers a decent albeit simplified explanation of the subject.
Wow cool, thanks for explaining!
As a longtime blender renderer and dabbler, this video was FANTASTIC in being visually captivating, well narrated, and hey, still learned a few new things, too! Subbed.
Thank you for the kind words, glad you liked the vid!
You dont know how much I appreciate you making me find out my render device setting was on NONE all this time, I knew it and I never noticed it, I feel so stupid but I'm glad I watched this today cause I have to render a lot of work tomorrow
Haha it happens to the best of us! Glad you got this sorted before doing all that rendering!
You missed something while talking about denoising. Under advanced settings in Cycles sampling panel there's animate seed option. That one needs to be disabled in order to avoid flickering between the frames. In fact, quite important. Also, I haven't seen anyone mention this but since 3.6 version, maybe even one or two versions earlier not sure, there's Turbotools add-on inside the Blender. That's accualy separate denoiser. If used properly can give even better results that Optix or OpenImageDenoiser.
That's not free though. I still use the ai denoiser from recycled. I've separated the addon from the special build and still jas way better results vs blender vanilla approach
Even better; apparently there is a temporal denoiser using Optix in Blender. But it's hidden somwhere under the developer extras. But yeah you're right on the seed for the noise threshold, good one!
@@KaizenTutorialsNever heard of it but I know EEVEE NEXT is a bit better about accumulating samples over frames which would be better for animations
@@KaizenTutorials the only place where I saw Optix temporal denoiser is inside the free Pidgeon Tools add-on. I've been playing with it for a while. Interesting results. Check that one. Cheers!
Nice! I'm not sure where it is. Southernshott has a video on it on his channel. @@_tomashh
Dude I've been rendering these animations for so long taking hours of my day just for a 5 second animation just to figure out I've been limiting my hardware and could've saved hundreds of hours just by clicking a button, my life is a lie
also thank you this video changed my life
damn that's rough! Happy this video helped you out tho
About using your CPU in addition to the GPU for rendering: It depends. Usually, it *should* be faster, but there also are a lot of cases/scenes where it actually slows down the rendering, in comparison to a GPU only render. Always test which is faster before rendering your animation
Yes! I agree; that goes for all these settings. Always test 1st!
On laptops, CPU+GPU is the worse option almost 100% of the time... Because either GPU or CPU alone can max out thermal/power constraints, it all boils down to how many samples per watt the system is producing.. By taking this some of this power budget away from more efficient GPU and giving it to far less efficient CPU will result in less performance.
It is always a pleasure to find new ways to speed up my workflow!
A few things to note: Make sure you are not thermal throttling when rendering with both the cpu and gpu (do not do this if you have a laptop it will reduce render times). Noise threshold can cause artifacts with denoising (especially with volumes). Experiment with the minimum samples to make sure you dont get these. Often times i find it best to just disable it entirely. Persistent data does not work with adaptive subdivision. I also find that keeping render times the same, rendering at a higher resolution results in more detail since the denoiser has more pixels to work with.
Thanks for the notes!
Thank you so much for these tips. I've been trying to render out a 1053 frame animation, but each frame was taking over a minute to render, even with the noise threshold set to 1 and samples reduced to 1024. I found the Optix option, played with the lighting settings and a few other things, and now, from my current test, each frame takes around 9-11 SECONDS depending on the complexity of the frame. Massive time saver ! ^_^
This Video helped me a lot for getting happier in Cycles. At least for me denoising settings were somewhat of a Problem. After turning off the denoise for the viewport i got much faster results in simple scenes.
The best results in terms of speed were achieved by turning Denoising off, but if you have the extra 2-5 Seconds in the viewport to denoise i recommend to set start sample to the last frame. This way not every sample gets an own denoising process, but only the last one does. So you wait less inbetween the actual Samples.
Haven't tried this in more complex scenes, but for hardsurface modeling this works like a charm.
Dont know if anyone said it before, but if you use a bigger format, like 4k instead of 1080 and use 1/4 of the samples the denoiser is able to mantain a lot more details and sharpness. You divide the samples by 4 because 4k is 4 times the area of 1080p. Max Hay uses this technique and downscales the render to 1080p afterwards. You don't actally have to change the resolution since changing the format % to 200% should be the same thing. Hope it helps someone! P.s. Please do a full video about temporal denoising techniques and fireflies/artifact removal :P Loved the video from Southernshot, but I was asking myself if there are any other workarounds (maybe even in AfterEffects or premiere)
Thank you for this! And yes maybe I will do a video on that in the future.
IT WORKED
the animation i have was a camera tracking one, so it took 3 minutes each frame, but now it only takes 30 seconds.
niceeeeee
good tips but no offense this video could be like 1/3rd the length that it is
it could be. It could probably be 1 min long, but there wouldn't be any fun or explanations in it. And I'm a strong believer of not just showing, but also telling!
I like the talking and animations. We are to used to fast videos and forgot how to enjoy some quality content ❤
@@lucas-correa watch it on 2x speed next time
Another tip: render from the command line. Some utilities online can do this, for those unfamiliar with shell scripting (like the Flamenco add-on, which I use), but this is another way of managing memory. This way, you don't have the full Blender UI running in the background while rendering - helpful for longer animations. A middle ground to this is to lock the Blender interface while rendering (meaning you can't use the UI while it's rendering) by going to the top menu -> Render -> Lock Interface.
Yeah i've heard it can shave off a couple of seconds, but I honestly couldn't get it to work myself. Then again I have 0 coding experience and understand nothing about it haha
@@KaizenTutorials that's fair. I have done a few tutorials in the past; maybe rendering from command line be my next one 😄
@@SpencerMagnusson I subscribed in case you do a tutorial in the future. Thanks for sharing your knowledge. The Blender community is so amazing.
@@studiojohnny thank you! I'm wrapping up some videos for BCON LA but I added this to my backlog to get into soon.
Here's my rig, it's an gaming rig tho my workstation is using Ryzen R9 5950 + RTX 3090, im sure you already know the performance in Nvidia side so here is the intel ARC GPU for you instead.
Intel i5-10400 + ARC A770 16GB
Render time :
GPU + CPU = 8m02s
GPU Only = 6m29s
note : since it's taking too long to render at 4000samples
I reduced to 2048 samples and noise threshold to 0.03
So clearly if you're CPU is slow it will actually hinder the performance.
Ah yeah so it seems, thanks a ton for sharing these numbers! As for the intel ARC gpu's; i'd hoped for them to be faster. Using NT 0.03 and 2048 samples takes my machine around 1 minute.
thatt persistent data is a world saver it saves so much time, thanks for sharing this amazing video
Glad I could help!
Yesterday I actually found out about these and my rendering time went from like 15 seconds to literally like 2 or 3 seconds (it was not a huge render thing but just a small one but doing this stuff helped 100%!)
Awesome, glad the vid helped!
omg I need to try this persistent data. why the hell is it not on by default??!
I believe because it takes up a lot of RAM, so that makes things a bit harder on low end PC's.
new to blender, was using maya and other soft, changing render settings sped up render about 10x. thank You
One thing I noticed is that some of my colleagues or students are using downloaded materials (from Blender Kit for instance) that are way too complex for no reason, like a glass with 3 different IORs and refraction values, or a shader with multiple 4D noise textures with the detail slider set to 15.
Also, one of my student had 8 hours render time, I was able to cut down the render time to 10 minutes just by some optimization, for instance some tiny objects had a PBR setup with 4096*4096 textures, completely useless !
In addition to what you said in the video, these are some important things to check :
- Polycount
- Texture sizes (also if a texture is 5000*5000px for instance, resize it to 4096*4096)
- Heavy procedural textures (the "detail" slider doesn't have to be set to 15 all the time)
- Volumetrics are killing your render times on Cycles (something like 10 times longer is normal)
Also, I noticed that Blender doesn't like to have too many objects (i'm talking about thousands of objects), if you can join them it can make a difference (or use instances if they are similar)
Thanks for the video !
Huge thanks! This cut my render time from 20 sec per frame to 4 sec!
Love it!
was looking for a rendering guide today and saw your old one with a link to this one, quite helpful thanks
Glad to hear and glad the link is doing it's job haha! ;-)
Biggest mistake I recently made was a total rookie move. Rendering a simple animation in cycles, but each frame was taking almost 2 minutes. Finally realized I just forgot to change the device from the default CPU to GPU. Went from about 2 minutes per frame to under 20 seconds! Sometimes the simplest setting makes the biggest difference.
We've all been there haha! Well done on figuring out the issue.
Not me waiting for my 350 frame animation to finish rendering while watching this video 😅. Usually, I significantly decrease the sampling and resolution, so these tips are a huge help
Aah dang haha, poor timing. But for next time you now know! ;-)
If youre using GPU, deactivate Tiling, Turn on Denoiser Optix or OpenImageDenoiser, low the samples to 12 and in output in resolution activate 200% With this you will denoise a bigger image that will have more details and you will get a better result with the denoiser because this one understand better the image. And deactivating the tiling, its because GPU handle better the whole image than tiling images. Doing test the tiling always got more time than the whole image.
Great tips, thanks!
I’ve seen a few people recommend that low samples high resolution trick and I’m a bit skeptical about it. At least in the scene I’m testing that has volumetrics, the results are quite a bit worse and has very noticeable blotchy denoising artefacts in dark areas. Not to mention the slightly slower render times even at such low sample counts and you’d be using exponentially more storage space if you keep the renders at that 200% scale. Doesn’t seem worth it to me.
To avoid running out of memory, I always save the project, restart the pc then render immediately after its on.
because restarting your pc will clear all ram and that's what we want.
That's a good tip too! Makes sure the system is fresh.
Great explanation I understood these concepts very well. Thanks, bud 😌
if only i had a gpu...
haha damn sorry bro
AFAIK, Tile size optimizations only works if you have more GPU cores than what cycles can use with parallel processing in a single tile. If you have equal to or less, then you are best off rendering 1 large tile. If you have more, your better off rendering a number of tiles that has the largest number of cores per tile simultaneously without going over. at least, that's how I understood it. I don't know how many cores cycles is optimized for. if it's 4 cores, and you're doing CPU rendering with a threadripper with like 60+ cores, multiple small tiles will be faster because it's rendering multiple tiles at a time, but AFAIK the average consumer-grade GPU can only render a single tile at a time, and reducing tile size and increasing the number of tiles, actually increases render time, rather than reduce it, because it's only 1 tile at a time, and switching tiles takes time to load new data to the GPU. I am NOT a programmer, or really any kind of expert, I just watch a lot of youtube videos, and tried to render on my laptop and this is what my experience of it points to. simultanious tiles = total # of cores MOD max cores per render.
So from what I've been seeingin the comments tiling is mostly useful to prevent VRAM issues and for CPU renders. Modern GPU's render faster on full tiles vs. small tilesizes.
@@KaizenTutorials is there a way to get a specific, optimal size for the GPU to take advantage of tiles, or is it always full image tile for GPUs now?
Thank you for this essential point, which raised existential questions! ....As persistent data!!!!!
Glad it was helpful, haha!
rtx4090 + this video = game changer! thank you for making this & for the render preset!
Glad to help! 🧡
Rendering my scenes on 2gb VRAM
Nvidia mx150 and other software says this GPU is out of date 😂
haha damn, but this will still help! I promise!
Thanks you so much, thanks to these tips I've gone from 3 hours to 11 seconds to reder a scene
That's amazing! So much time saved.
There is mistake in the description of how Light Paths works. Transparent is not responsible for bounces in the glass, it works in bounces of the alpha texture in the shader (ex. leaves in trees with alpha mask)
You're right! But that also affects glass believe; atleast it's transparent component. which you can see in the render as using these bounces did make it see-through again. But for the light refraction inside of glass you'll need other bounces!
@@KaizenTutorials You can doublecheck it with array of glass planes on front of Camera. If you add alpha weight to you glass shader, transparent bounces will work
This was so helpful 🙏
I was told several times earlier to deactivate CPU from Optix, because it "gets in the way", though both my processor and graphic card are not as good as yours.
Also the scene is amazing!
Ah yeah it seems to depend on the hardware honestly. On my current and previous setup (3070 and ryzen 3700X) using both was faster though! And thanks, appreciate it!
@@KaizenTutorials I`ve got 3060 TI and Ryzen 5 2500x. guess I need to do some tests
Love the tips shared and the delicious salad! Appreciate the sense of humor while learning! 🤣 U R Z Best man!!!🤩🤩 Luv it! 😍
Ah thanks a lot!
That thumbnail🔥🔥
Hey ! Thanx ! Love your videos, they are helping me so much and I also like the style.
Don't mind what some individuals with ridiculously empty calendars say in the comments.
Thank you for the kind words!
Your videos are great, informative and entertaining. Thanks bro
Glad you like them, thanks!
Dude saved thousands of lifes❤
the thing I was trying to render went from 8 hours to 3 hours 😭 WAYYYYY BETTER
That Witcher and Oppenheimer soundtrack on your Spotify...🔥👌🏻
Soooo good!
Nice video!
Remember that Salad using you GPU is not free, electricity cost money.
That's true ofcourse!
I always crank the render settings to 1024 tile size and 2048 samples to then denoise outside outside Blender or Maya.
I bought a 3090 four days ago and after coming from a 1060, 2060, 4060... it's a breeze. Something funny is that the 3090 is much more accurate at tracing than all of the GPU's i used so denoising is very easy even with few samples.
honestly before watching this video I was thinking of just sending my projects to render farms becuase my blender would crash immediately, i am used to just modeling and rendering a single frame for small simple assets so I dont save my project till the very end, my mistake became obvious very quickly when i noticed my whole project was gone, including all my assets I made in the same project, dont do what i do...
Ah damn, yeah always make sure to save!
Appreciate the tutorial! Good tips in this one.
Thank you!
merci pour vos conseils avisés.
De rien.
For tile size:
It did massively help for me on the CPU (especially in multi threaded systems), but in my experience, it's comparatively slower on the GPU
Thanks for your input!
I love those fast renders. At one point I didn’t know exactly what I was doing and I would render a billion frames at an overly large size. It would take an hour to render when really I needed one square frame. 😅
HAha yeah I feel like we've all been tehre!
Step 3 is completely wrong
Optix is fast af, but Openimage is huge WAAAAY more better quality.
And to get rid of "Denoising artifacts" on animation you just need to set noise seed (under advanced tab) any number than 0. So the noise pattern would be the same every frame and there will be no denoise artifacts on animation and the animation would be smooth ASF
In my tests Optix performed about 90% the same as OID for stills, with OID being the clear winner, for sure. But for animation I don't like it at all. The seed is a great tip, but apparently there's even a full blown temporal denoiser available in the developer extra's right in Blender!
Honestly I barely use the noise threshold because I like the more obvious result of just increasing the number until it's too slow. I feel like the not noisy areas don't take long to render anyway
If you have the Noise threshold off all areas will take exactly the same amount of time since Blender will render all samples for every pixel.
I don't wanna be a shill, but there is a addon I use to use a hell of a lot called "Super Image Denoiser" which does have a temporal denoising setting. Back when I used it it could do a pretty amazing job at denoising scenes, but nowadays I'm fairly sure it's broken as it's free version (I think) has stopped development and is now been put into a pretty nifty addon tool kit
No worries, thank for mentioning this improvement for people.
Helped me a lot! you will help me save minutes/hours on my projects
Awesome!
another way to speed up renders is like they use in vidddygame engines is to compile all them texture node to trim sheets and hero textures , that way the compile time is in milliseconds , also like show'd with single color backgrounds is to use the just a diffuse node not a whole procedural node when needed , also a make pass renders for each of the diffuse , glossy and transmission and then can use a compositor later to fix the colors and intensity as well as speeding that by dividing scene into collections for foreground , mids , background and holdouts
good tips!
@@KaizenTutorials can u make a video on what he mentioned, video would bang
Thanks for this! I had everything set like you did minus making both CPU and GPU run at the same time. Changed that one setting and my 240 frame animation went from a 45 second render per frame, to a 5 second render per frame
Aah that's so nice! Glad this helped.
Great video, great explanation, straight to the point, and really made a difference, thnx bro
Glad to hear it!
I was alway using OID for animation because I heard in the past that it was better for animation. I Guess it's not the case anymore, gonna try optix Denoise now.
There’s apparently also a secret temporal denoiser! Southershotty has a video on it. It also uses Optix, but is hidden behind developer extras.
Quick heads up about 2:53 : Render times can still be slower if you have both CPU and GPU enabled. I tested it in a scene I was working on and it was quite a bit slower (I'm using OptiX).
Ah yeah that's true. It's really dependent on hardware, scene and setup.
Thx bro !
WHY DIDN’T I SEE THIS EARLIER I AM SITTING IN FRONT OF MY LAPTOP WITH 2 DAY RENDER RUNNING OMG
Damnnnn 💀💀
Good stuff
Cheked to see the pc specs in the description only to find out that i literally have the same pc, only a different AiO cooler. Noice!
Awesome haha! 👏🏻👏🏻
Got the rendering time to 59,75sec with rx 6600, ryzen 5 3600 and 16gt.
Gpu only, no tiling, no noise threshold sample 64. OpenImagineDenoise accurate, seed 1, max bounses 5,4,4,4,0,8, Fast GI, Simplify 1 max divisions 12 & texture limit 8k. Use spatials, persistent data.
Great! what was it before optimization?
Dude is this true ?
“KAIZEN™ means improvement. Moreover, it means continuing improvement in personal life, home life, social life, and working life. When applied to the workplace KAIZEN™ means continuing improvement involving everyone - managers and workers alike.”
Yes! It's the very reason I chose this name for the channel and my business. :-D
@@KaizenTutorials to stay motivated its amazing mate !💗
Thumbnail is very eye-catching, but it took me a minute to understand it. I think switching the places of the button and text would be much better!
Ok! Thanks for the input. 🙏🏻
Enabling OptiX only for my GPU, and not for both CPU+GPU, improved my render times, from 12min to 9min
And my CPU is not at max usage now
(7800X3D and RTX 4080)
Edit: Well Blender disable your CPU for OptiX if you have a gpu enabled, at next startup
I just love this video, amazing
Thanks love to hear it!
Kaizen God.
to tile sizes: by setting your tile size to the amount of bits in your GPU it can shave off multiple seconds (if you don't know your bit-count, use 256) i don't know much about this but it pretty much just renders the single squares and if the GPU doesn't need to cache anything it'll be faster i guess
Awesome! Thanks for sharing.
you are the best ... life saver
Amazing Videoo Fun to watch !!
Thanks a lot!
Very good explanation!
Thanks!
You missed one of the crucial settings that gives 50 percent more boost 😊 just un check the tree option and see miracles 🎉
Light tree under lights ❤
Never heard of this but will check it out! Thanks!
@@KaizenTutorials thank me later 🤗
This video is a gem
Thanks so much
Wow, thanks for the kind words!
Use small tile size can help reduce VRAM in my experience
So i've heard yeah!
Duuuuuuuuude what an ammmazing video!! Thank you soooo much
Glad you liked it and thanks, appreciate it!
THANK YOU for your excellent suggestions. You saved me a MASSIVE amount of render time.
Glad I could help!
u know i never played with the noise threshold before this . . . u have saved so much time on renders . . . from 5 min to under 1 . . . with openimage its still fairly good unless zoomed in on reflections
Awesome, glad I could help you!
faster isn't necessarily better however and you can then find yourself fighting artifacts you don't understand as a beginner because you followed a make blender render faster tutorial. I just wish these videos had a bit of disclaimer in that regard.
I think I mention that several times. Speed comes at the cost of quality.
In my experience, tile size just puts less strain on the hardware. I kept the tile size the same while I was learning blender, but kept getting crashed when trying to render multiple donuts. Now Don crash yet
damn this tips made my renders waaaaaaaaaaaaaaayyyyy faster
Glad to hear it!
The timing of this video is no coincidence.
Good luck to all of you who are rushing to crank out a render before the Lunar Landscapes deadline! 😬
Haha it is not indeed! Let's hope this saves some people and let's them submit on time ;-)
Thank you
Rendertime with Blender 4.0.2 for the original scene from the link in the video-description, frame 1, no changes made: 05:25:27 minutes!!! Hardware: AMD 3970X Threadripper (32 Cores); ASUS ROG-STRIX RTX-4090 OC; 128 GB RAM; 2 TB Samsung SSD 970 Pro; OptiX, Cycles CPU+GPU-rendering; But turning down the samples and the threshold will make it faster for sure.
Rendertime with changed parameters: Noise Threshold 0.05; Bounces: 8, 4, 4 ,8 ,0 , 8; Persistent Data; Max Samples 2048 equals 00:31:21 minutes!!!
you should turn off the CPU and only use the GPU, and the default sample rate of 4096 is unnecessary high if you try to render a single image. What do you mean with original scene?
@@MrDragon005That's what I never got a clear answer for at discord-channels: Is it better to use CPU AND GPU, if having a "not sooo slow one", like me, or use GPU-only? Please explain your opinion. :) "Original Scene" means, that I have not changed anything, in the scene, before hitting "F-12". :D Rendertime is 00:28:14 minutes if using GPU-only!!! WHY????
That’s awesome!
Both is better in my testing both in my current setup as my old one with a 3070 and ryzen 3700x.
Using 4096 is fine because changing the noise thresholds prevents unnecessary sample renders. So you can leave it up. As I show even setting the samples to 20K doesn’t matter because the noise threshold never let’s you get there!
I think your hardware setup is missing from the description
You're right! Just added it in, my bad!
bro Love from Heart
This is great!
Glad to hear it, thanks!
Surprised you havent mentioned scrambling distance. Its an absolute lifesaver on cityscale+ scenes
I don't even know what it is! Care to explain?
Thankyou this had everything I was searching for and more it really helped me
Awesome, glad to hear it!
Great vid, i learned a lot
Thank you!
ANOTHER TIPS:
u can use ai upscaler (CupCat AI upscaler) and render with half the resolution desired.
u can use ai frame interpolation (FlowFrames) and render with half the number of frames
Yep, good tips! They tend to work a lot better for projects with slow, predictable motion, but if you have one of those it will save you hours and hours!
Once again we see that blenders default settings are for some reason set to the worst possible thing.
Haha true! I think they're just there to be modified by everyone based on their projects.