00:00Look at this AI video
00:02I generated this using SDXL 2.3 in ComfyUI and it only took 8 GB of VRAM
00:09Running SDXL 2.3 usually causes an out of memory crash on an 8 GB GPU
00:15But today, I'm going to show you the exact patches and settings to run it safely
00:20>> [clears throat] >> And to prove that it works, we are going to watch the performance monitor live so you can see the VRAM stays safely under 8 GB
00:31Now, this workflow is very simple to use
00:32I did not make it complicated
00:36This time, I kept it simple
00:39So, let's set it up
00:39First, we need to add a few simple commands to run ComfyUI in aggressive memory saving mode
00:48Normally, you just type python main.py to start, but today, we are going to add low VRAM
00:56Next, we need to add reserve VRAM 102
01:00And what does this do? It keeps 1 to 2 GBs of memory free for Windows
01:06Therefore, your computer will not crash
01:10For my test today, my GPU has 32 GB
01:15Therefore, I am typing reserve VRAM 24 to leave exactly 8 GB for ComfyUI to use
01:23Next, we need to fix the cache
01:26Add cache LRU 10 to reduce your cache memory usage
01:32Finally, add preview method TAESD to create a very lightweight image preview
01:40But if you still get an out of memory error later, just change this to preview method none to save even more memory
01:45If you use the portable version of ComfyUI, do not worry
01:53Just right click on your run NVIDIA GPU bat file, click edit and paste these exact same commands after your main.py
02:02Now, let's look at the models
02:05First, I used a distilled 1.1 Q4 KM model
02:10This saves a massive amount of VRAM because we do not need to load an extra Laura file
02:16Next, we move to the dual clip loader
02:20I load the Gemma 3 GGUF file
02:23And why do I specifically use the GGUF version? Because this compresses a massive AI model down to just a 6.14 GBs
02:35This is exactly how we force it to fit inside low VRAM
02:41After that, I load the LTX 2.3 text projection file, which is about 2.31 GBs
02:47All together, these models add up to around 23 GBs
02:53I also added a low VRAM patches group to this workflow
02:56We take our loaded model and pass it through Sage attention, memory efficient attention, and chunk feed forward patches
03:05These nodes are essential because they reduce your peak memory usage while the video is actively generating
03:14We saved our memory, but we also need to stop the AI from generating bad anatomy
03:21Therefore, we use the LTX nag node
03:25And what does this do? It makes your negative prompt much stronger
03:31It forces the AI to listen when you say no bad hands, no weird text, and no artifacts
03:38This one simple node makes your final video look better
03:42Next, let's look at the prompt section
03:46I added a bypasser group here
03:49You can write your normal prompt manually, but what if you want to use a prompt relay? Therefore, I added a prompt relay option
03:59You can easily switch between your manual prompt and the prompt relay, depending on what you want to do
04:06After your prompt is ready, we need to set the video size
04:10Look at my note right here for the recommended resolutions
04:15I know some people use a custom resolution node, but every extra node waste your limited VRAM
04:22Therefore, I keep it simple and writes the value in plain text
04:27I just type the numbers manually
04:30For an 8GB GPU, I highly recommend starting with a resolution of 480 by 832
04:37I set it exactly 10 seconds at 24 frames per second
04:44You can lower this value later if you need to save even more VRAM
04:49Now, here is my trick
04:51I divided the video generation into two passes
04:55The first pass does the first six steps
04:58After that, the second pass does two more steps
05:02And why do we do this? Because if you are on low VRAM, you want to make sure the first pass works safely before moving forward
05:13Therefore, I highly recommend you bypass the second pass and the upscaler first
05:20So, let's test the first pass live
05:23I'll hit run and open my performance monitor
05:30Watch the graph closely
05:32Loading the text encoder model only takes around 3.7 GB of VRAM
05:38And when it goes to the prompt section, it takes around 5.6 GB
05:44Now, it's on the first pass and the VRAM spikes up
05:48Right now, it takes exactly 8.1 GB to generate this video
05:54It only takes about 44 seconds to generate the first six steps
05:59After it finishes, you can see my VRAM drops back down to around 3.6 GB
06:06We got a great first pass, but we need to finish the video
06:11Therefore, I'll enable the second pass and hit run again
06:18Because we used the correct cache setting earlier, it does not start from the beginning
06:24It starts exactly where the first pass ended
06:27In the second pass, it takes around 8 GB of VRAM again
06:33These last two steps only take 15 seconds, meaning the full video takes exactly 1 minute to complete
06:44Now, we have a great video, but what if you want to upscale it to make it look even better? So, let me show you what happens when I enable the built-in upscale section
06:57Watch the performance monitor
06:59The VRAM passes 8 GB and spikes all the way up to around 14 GB
07:05It will still run on an 8 GB GPU and will not fail
07:10But because it uses so much memory, it will take a very long time to finish your video
07:16Therefore, I use a much better method
07:20Instead of the built-in upscaler, download your generated video first, then load that video back into ComfyUI and use the RTX video super resolution node
07:33And what does this do? It upscales your video safely
07:38You can even set it to ultra quality, and it only takes around 3 to 4 GBs of VRAM
07:45It takes about 30 seconds to generate and it's much better for low VRAM computers because you do not have to wait a long time for a 14 GB memory spike
07:57Now we saved our memory and upscaled the video safely
07:59Now let's look at the more results you can get with this workflow
08:05Here is the first example
08:07A girl is singing a song
08:10The camera movement is highly realistic and the quality is amazing
08:17I generated this entire scene in around 1 minute
08:21But what about complex physics? But what about complex physics? So here is the another scene I created
08:29An old man is sitting and smoking while a train passes through the station
08:35Watch the physics closely
08:35When the train passes, the wind actually pushes his smoke in that direction
08:43There is a very small physics issue if you look closely, but to the naked eye, the overall result looks fantastic
08:51So that's all for today's video
08:51If this 8 GB ComfyUI setup helped you, hit subscribe so you don't miss my next AI tutorial
08:59And if you want to learn how to make even better AI videos, click the video on the screen right now where I show you my next secret
09:05See you there
09:05Bye-bye.