跪拜 Guibai
← Back to the summary

Running ComfyUI on an RTX 5060 Ti: FP4 Makes the Budget Card Competitive Again

by 雪隐_上班了 from juejin.cn/user/143341… Sharing and aggregation are welcome, but full-text reprinting is not necessary. Please respect copyright. The circle is only this big. If you need it urgently, you can contact me for authorization.

Column Purpose: Use my RTX 5060 Ti 16G + 64GB RAM "budget battle wagon" to do some novel things that justify the electricity bill.

In this chapter, we'll make the graphics card "draw" — and not just kindergarten-level stick figures, but the kind of stuff that lets you show off on social media.

Foreword: Why Did I Delay This Update for So Long?

Actually, I hesitated for a long time about whether to write this chapter. There are three reasons:

  1. ComfyUI has too much content — Different people use it for different things: some generate images, some generate videos, some do cutouts, some work on character consistency, some do style transfer... Each direction could be a whole series, and just the table of contents could fill half a book.
  2. It updates too fast — Today you just learned a workflow, tomorrow a new model comes out, and the whole workflow changes. The speed of writing tutorials can never catch up with the version numbers. I might finish writing, and before I even publish it, the plugins have already updated.
  3. I'm afraid of being criticized — Because my way of playing might be different from yours. If I write too briefly, you say "That's it?"; if I write too detailed, you say "Padding the word count." It's a dilemma.

But later I figured it out: I don't care about you, I'll write my thing, you read yours. Anyway, this column is called "Playing with AI on a Personal Computer," not "ComfyUI from Beginner to Mastery." Sharing my personal experience and the pitfalls I stepped into is enough.

So, here comes this article — an ordinary person's mental journey of playing with ComfyUI on a 5060 Ti, including three times I almost uninstalled the software.

1. Getting Started: I Recommend "Copying Homework"

If you ask me: "How do I get started with ComfyUI?"

My answer is just one sentence: Go find Qiuye's (Autumn Leaf) integrated package.

Portal: Qiuye's Bilibili Space

Download, unzip, double-click to run, and you have a working ComfyUI.

Why don't I teach you how to install it from scratch? Because I'm not here to be a teacher, I'm here to be a porter. Qiuye has already filled in 90% of the pitfalls for you. If you insist on stepping into them yourself, that's your freedom, but I won't accompany you.

2. A Few "Mods" I Made

Qiuye's integrated package is very useful, but I'm a programmer who can't keep my hands still, so I changed a few things:

1. Deleted the built-in Python environment and replaced it with Conda

The integrated package came with a Python environment, but my OCD kicked in — I already had a Conda on my computer and didn't want to raise another Python orphan. So I deleted the built-in Python and created a new environment with Conda.

conda create -n comfyui python=3.10
conda activate comfyui
pip install -r requirements.txt

Why choose Python 3.10?

Everyone online says 3.10 has the best compatibility; 3.11 and 3.12 will cause errors with some plugins. Although 3.13 is already out, I'm not here to be a lab rat. Use whichever version is the most stable, same logic as blind dating — don't look for the prettiest, look for the most reliable.

2. Got a problem? Leave it to AI!

When I first opened ComfyUI and saw that interface looking like a circuit diagram, I was petrified for three seconds on the spot.

But later I discovered a truth: 95% of problems, AI can solve for you.

The remaining 5% are "incompatibility" problems — the kind that even AI can't solve, usually for two reasons: wrong Python version or mismatched CUDA version. At that point, don't struggle, just switch environments and start over.

Someone asked me: "Why don't you learn it yourself?" I answer: "If AI can solve a problem, don't waste your own brain cells."

3. Advanced: A "Stocking Up" Guide for Models and Plugins

With the basic environment ready, the next step is to add things in. My approach is very simple:

Browse Hugging Face. Download any model that looks interesting. Throw it into ComfyUI. If it doesn't run, delete it. If it runs, screenshot it and post it on social media.

It's that crude.

Models I've Tried (Incomplete Statistics)

The "Evolution" of Hard Drives

Another side effect of playing with ComfyUI is: Hard drive space is never enough.

A single model is easily over a dozen GB, a plugin comes with a bunch of dependencies, a base image iterates repeatedly and eats up cache. ComfyUI consumes hard drive space faster than my year-long shopping cart.

4. Quitting Phase: I Almost Stopped Playing

Later on, I started trying 1-minute video generation and character consistency maintenance — making the same character look identical across multiple scenes.

That feeling was like doing a jigsaw puzzle; very fulfilling when it fits, wanting to smash the computer when it doesn't.

Until I saw closed-source models.

Nano Banana 2, Sora 2, Seedance... One by one, closed-source models descended like divine soldiers, the video quality they generated directly outclassed the results I had tinkered with for several all-nighters in ComfyUI.

And they had no requirements for your machine — click a button, the cloud runs it for you, you just need to pay.

At that moment, I sat in front of my computer, looking at the 720P video my 5060 Ti had chugged out, then looking at Sora's 4K demo on my phone, and fell into deep thought:

"Is there any point... in me messing with all this?"

Then I closed ComfyUI and played games for a week.

5. Returning Phase: FP4 Brought Me Back

Just when I was about to forget ComfyUI in a corner of my hard drive, one day I was browsing the news and saw NVIDIA promoting NVFP4 (4-bit floating-point format).

Simply put: FP4 can shrink model size by 4-5 times, speed it up several times, and is only supported on RTX 50 series graphics cards.

I glanced at my own 5060 Ti, and a thought popped into my head:

"This budget card of mine... is actually on the new architecture?"

Before, running Flux was like bathing an elephant — not enough water (VRAM) budget, time-consuming and laborious. After switching to the FP4 version of the Flux model, the 5060 Ti 16G actually ran it steadily, and at a speed I never dared to imagine before.

Conditions:

I tried it, the quality didn't degrade much, but the speed increased several times — that feeling is like you spent a few thousand bucks on an old car, and suddenly someone tells you "change to this barrel of oil, and it'll run like a Ferrari."

The picture below is the running effect (screenshot kept, just feel the vibe):

flux2

LTX 2.3 can also run, speed is touching (the good kind of touching):

ltx2.3

So, the current mood of 5080 and 5090 users:

"The stuff we've been running for over half a year, you 5060 Ti guys can run it now too? Then the extra money I spent is..."

Don't be sad, your image quality is better. We just solved the problem of 'can it run', you solved the problem of 'how well does it run'. One is in the sky, one is on the ground, the gap is still there.

6. Hands-On: Snag a New Model from Hugging Face

All talk and no practice is a fake show. I'll take a new model I recently saw on Hugging Face as an example, to show you the actual operation process.

Target: conradlocke/krea2-identity-edit

This model can do one thing: Maintain character consistency while modifying the scene / merging multiple characters. In plain terms: you give it a picture of Leonardo DiCaprio's face, and it can put him on the moon, in ancient Rome, on a cyberpunk street, and he still looks like Leonardo DiCaprio.

Address: https://huggingface.co/conradlocke/krea2-identity-edit

Step 1: Look at the Page

The model page usually provides ComfyUI workflow hints, click directly in:

Workflow hint

Step 2: Download the Workflow JSON

Enter the page to download the workflow file, drag it into ComfyUI:

Drag in workflow

Step 3: Handle Missing Nodes

You'll likely see a screen like this — a bunch of red errors, like a flatline on an ECG:

Missing nodes

Don't panic. Click 'Install Missing Nodes':

Install Missing Nodes

ComfyUI Manager will automatically handle most dependencies for you, a small portion needs to be manually found on GitHub. But don't worry, there's an 80% chance AI can tell you how to install it.

Step 4: Download the Model

Go back to the model page, download the relevant .safetensors files, and put them into the corresponding models folders:

I took a shortcut: Swapped the main model for the FP4 version krea2_turbo_nvfp4.safetensors, so it runs faster on the 5060 Ti.

Replace FP4 model

Step 5: Click "Queue Prompt"

Wait for the progress bar to finish, and whatever picture comes out — no matter the effect, screenshot it first and post it on social media, with the title "Briefly tried out a new ComfyUI model."

Whether it looks good or not isn't important, what's important is you got it running. Others don't know how many times you ran it to get this one picture.

7. Practical Advice: Don't Run Image Generation When It's Hot

Finally, a lesson learned in blood and tears for everyone:

It's summer now, indoor temperature is 30 degrees. I ran a video generation, and the graphics card temperature directly shot up to 77 degrees.

What's the concept of 77 degrees?

Although 77 degrees is within the "normal range" for a graphics card, seeing that number just makes me uncomfortable, just like seeing my bank account balance — the number itself isn't a problem, but it just looks painful.

Suggestions:

An old buddy once asked me: "Why does my computer always crash when I run image generation in the summer?" I said: "Touch your case, can you fry an egg on it?" He said: "Actually, yes." I said: "Then who are you blaming?"

8. Summary: I'm Still Playing, But Not So Obsessed Anymore

ComfyUI, this thing, is not hard to start, not easy to master, updates too fast, and makes you want to quit anytime.

My mentality changed like this:

  1. Just started: "Wow, so powerful, I want to learn!"
  2. Played for a month: "I'm practically an artist!"
  3. Saw closed-source models: "Is there any point in me messing with this..."
  4. Discovered FP4: "Hey, I can play again!"
  5. Now: "Play when I have time, leave it when I don't, go with the flow."

I won't stay up late debugging workflows like before. But I won't uninstall it either — who knows, maybe someday there'll be a new model, new format, new surprise?

This is probably the normal state of "playing with AI personally": No need to keep up with the latest, just seek to have fun playing.

[Resources mentioned in this chapter]

Written at the End

This article is a sharing of my personal experience, not an authoritative tutorial. If you find it useful, likes, comments, and forwards are all welcome. If you think it's poorly written... well, there's nothing I can do, this is just my level.

By the way: If you also want to run ComfyUI on a 5060 Ti, feel free to exchange tips. Let's research together how to squeeze the last drop of performance out of this "entry-level card."

Thank you everyone 🙏 May your generated images never have distorted faces, and your graphics card temperature never exceed 70.