Hello, I'm Deno. This video is about optimizing MiniMax. We'll make videos much faster under the same conditions and use a new custom node I built so you can generate longer videos at higher resolutions.
I'll share what I've learned through my research. Let me start by showing you how much faster it gets. In my previous video, I applied a technology called SageAttention.
I mentioned that a 15-second, 1-megapixel video took roughly 900 seconds to generate. That's 15 minutes. I recently posted another video as a YouTube Short, and that one took 17 minutes.
So even with an RTX PRO 6000 GPU, Even a 15-second, 1-megapixel video takes 15 to 17 minutes, and it is still only 1 megapixel. The image and video quality weren't exceptional, so the result wasn't very satisfying. Then Sol-Attn recently started gaining attention, and implementations began appearing.
With no optimizations at all, each video takes around 23 to 30 minutes. With SageAttention, it took 17 minutes. Add Sol-Attn, and that drops to about 10 minutes.
I was very happy with it, with almost no loss in quality. Then I added a Fused Modulation node and saw a small further speedup with no loss in quality. So I began using SageAttention, Sol-Attn, and Fused Modulation together by default.
That brought the time down to around 9 to 9. 5 minutes. I kept using that setup until a member in our Discord shared what you're seeing now: a newly released Latent Upscale model.
Their test results looked excellent. So I downloaded, installed, and tested it myself. Looking more closely reminded me of when MiniMax first came out.
I once mentioned that I had spent several days trying to upscale it before eventually giving up. The barrier was that I'd need to train a LoRA or train a separate model on the latent space. So I moved on.
But when I checked this repository, it said they had trained a dedicated upscaling model on roughly 80,000 samples. The dataset was substantial, and it was clear this was a serious effort. My tests were extremely impressive.
I spent about a week testing before sharing it today. Normally, when generating with MiniMax without a LoRA, we simply run it for 20 steps. With LTX, we used an 8 + 3 step method: build the draft in the first stage, then finish with upscaling in the second.
The core idea here is the same as LTX. Generate the initial pass first, then double the size for the later stage and upscale it, then use one more pass to produce the final result. My recommended setting is 15 low-resolution steps with the MiniMax model, followed by the remaining 5 steps at high resolution.
To put it simply, I used to generate videos like this at 1 megapixel for all 20 steps. So it was 20 steps at 1 megapixel. Now that Latent Upscale works well, we can generate the first 15 steps at 0.
25 megapixels, then finish only the last 5 steps at 1 megapixel, for a total of 20 steps. So 75% of the process runs quickly at low resolution, Only the final segment is refined at high resolution, dramatically reducing the total generation time. Now let me show you how much time this approach saves.
SageAttention alone took 17 minutes. Adding Sol-Attn brought it down to 10 minutes. The video you're watching now uses Latent Upscale, and it took 256 seconds.
That's about 4 minutes and 16 seconds. Going back to the baseline, with no optimizations, a stock setup took 25 to 30 minutes, so let's conservatively call it 25 minutes. With the optimized workflow, the same quality, resolution, and duration take 4 minutes and 16 seconds.
That's a time saving of roughly 21 minutes. The difference is huge. I also researched a setup for those of you using mainstream graphics cards.
The test system used an RTX 3060 Ti with 8GB of VRAM. Let me play the video. This video was generated on the RTX 3060 Ti, and each run took exactly five minutes.
It produced an eight-second clip in five minutes, so the sub-PC test is now complete as well. Another question many of you may have is: "Why not use Turbo LoRA? " "Or Spectrum, or a cache-related node?
" Regardless of VRAM, I generally avoid optimizations that are structurally prone to severe quality loss. I care less about whether it merely runs and more about whether it's actually usable. New optimization workflows, proposals, and opinions appear every day in the open-source community.
But this is the workflow I've decided to use for now, after extensive testing and careful consideration. So please keep that in mind. I've explained the reasoning, so now I'll briefly show you how to use the workflow.
Applying this workflow is quite complex, so I designed it around using an agent. At the beginning is my preferred system prompt, refined through extensive testing. It tells the agent how to write the prompt.
Second, I documented how the workflow is designed and how it operates, so the agent can understand the entire setup. Finally, this workflow requires several different custom nodes. It's not enough to install a few custom nodes; you also need the correct ComfyUI version, along with several Python wheels.
The setup is quite complicated. The setup is too complex for beginners to do manually, so I documented how to build the environment. You don't need to read everything and install every component yourself.
Just show it to an agent and say, "Get this running. " It can set everything up automatically. I previously posted about the node in the center, explaining that I'd created a new node.
Normally, when ComfyUI runs, the text encoder loads and analyzes the user's prompt. It passes that information forward to the sampler, then decodes the result into a video after sampling. But I learned that ComfyUI's memory management isn't completely optimized.
As you can see here, even well into sampling, part of the text encoder still remains in VRAM. Ideally, only the diffusion model needed for sampling would occupy the VRAM. But part of the text encoder remains loaded, preventing efficient VRAM usage.
Insert a node that fully unloads the text encoder, you can potentially make 100% of the VRAM available while sampling. That's what this node does. After building it, I shared it on Reddit.
I told people it was useful and worth trying, but the comments were surprisingly negative. Without even trying it, people said things like, "Memory management is sensitive, so be careful. " Others said they'd generated a few images with Krea 2 and saw no difference.
The reaction was very negative. I found it absurd, so I said, "Fine, don't use it," and removed the post shortly afterward. The comments were all wrong, and almost nobody understood what the node did.
It did receive plenty of upvotes, but Reddit used to have more experts and experienced community members. Too many people there now seem poorly informed, so I rarely participate anymore. I mention this because Reddit used to be a community with lots of high-quality information.
Now it has grown so much that many active users seem to lack even basic knowledge. So when looking for new information, I think the official ComfyUI community may serve you better than the Stable Diffusion community, or you could follow developers like Kijai directly on GitHub. I think those sources will be more useful going forward, so I wanted to mention that.
Two users left great comments. One said, "I expected no real difference, but this is incredible. " I was glad they liked it.
Another said, "This is insane. It's so much faster," with a heart. It has no apparent side effects, and it works in any workflow, not just MiniMax.
So when asking agents to build a workflow, instruct them to always include this node. Then, even without mentioning it every time, you can create higher-resolution, longer videos. So keep that in mind.
I'll upload the workflow to Civitai and leave a link. When you use the workflow, open it in Codex or Claude, then drag and drop it, or copy and paste it, and say, "Set up everything needed to make it run. " The agent will apply it right away.
That's all for today's video. Thanks for watching.
