Fulfill your dirty fantasies with Comfy UI

Yo! HoangYell’s out — and in his place, you’ve got his top disciple, YellFilmHell. 🎬 Welcome to your first lesson in the dark arts. Complete it, and boom — instant graduation. You might not get how it works, but who cares? It does. Just enough to do the bad stuff. Hehe 😈

Oh, and I see your IP. If you’re a kid, close this blog now — or I’ll call your parents. 👮‍♂️😈

Kids gone? Good. The rest of you, follow me.

We’ll do this in three passes: pick ComfyUI, survive the hardware gate, then walk the full beginner path from comfy to your first saved PNG. The animated map lands right after the install commands—same story, one spine.

Let’s begin. 🚀

TL;DR

What it is: Local node-based image generator—ComfyUI runs on your GPU, you own the workflow.
Why it matters: No cloud filter nagging you when ChatGPT or Midjourney says no—full control on your machine.
Best for: Adults with an NVIDIA GPU who want uncensored, repeatable, steal-able Civitai Red workflows.
Key spine: Load Checkpoint → CLIP Text Encode → KSampler → VAE Decode → Save Image (the scene animates this).
Where to browse NSFW models/workflows: Civitai Red — civitai.com is SFW-only since April 2026.

Repository: comfyanonymous/ComfyUI

Beginner Map

Do not read this like a random tutorial dump. Read it as one loop: install → load weights → wire nodes → prompt → Queue Prompt → PNG in outputs/.

Pass 1: Why ComfyUI beats dead A1111 forks (modular graph, not mystery menus).
Pass 2: Hardware gate—VRAM decides full Flux vs fp8 (see the Pick checkpoint beat).
Pass 3: Watch the eight-beat scene after setup; memorize the Generate spine.
Pass 4: Hell Mode for real weights, Prompts for CLIP, LoRA for spice—then run the exercise below.

Term	Question it answers
Checkpoint	Which `.safetensors` file is the brain?
CLIP Text Encode	Where does my prompt text go?
KSampler	Where do I click Queue Prompt?
Workflow	Why can I drag a Civitai Red PNG onto the canvas?
LoRA	How do I add style without swapping the whole model?
VRAM	Why did full Flux lag but fp8 worked?

First Practical Exercise

After comfy launches in your browser:

Download any Civitai Red PNG that includes workflow metadata (or use ComfyUI’s default graph).
Drag the PNG onto the canvas—watch the Import workflow beat happen for real.
Write one short prompt in CLIP Text Encode, click Queue Prompt, wait for the run to finish.
Open ComfyUI/output/ (or your configured folder) and confirm a new image landed.

If those four steps work, the rest of this post is seasoning—not survival.

Why ComfyUI? 🤔

There are 3 popular local text-to-image client tools right now: Automatic1111, Forge, and ComfyUI.

Tool	Status / Description
A1111	Basically dead — no new commits since July 27, 2024
Forge	Slightly improved fork of A1111 — better, but still limited
ComfyUI	The best: most flexible, most powerful, with a transparent, modular user experience

ComfyUI wins because you see the wiring—nodes and arrows, not mystery menus. Every section below hangs off that idea. The scene after setup is the same graph, animated.

What Your Computer Needs 🖥️

Component	What You Need to Slay
GPU	NVIDIA RTX 3060 12 GB or beefier (8 GB VRAM minimum). Want Flux.1-dev or SDXL to sing? Aim for16–24 GB VRAM, you baller.
RAM	16 GB to keep it smooth; 32 GB if you’re multitasking like a caffeinated octopus.
CPU	6-core modern chip—CPU’s just the sidekick, GPU’s the superhero here.
Storage	SSD with 50 GB for models + 20 GB for your spicy output stash.

Note: These are the minimum specs. If your PC can’t handle it, just like the kids who got kicked out of class — you’re out too. Go build a new rig before you come back. Check out my blog for a simple PC build guide: 👉 How I Built My Dream PC from Shopee 🛠️

Hardware cleared? Good. Act I is terminal work—venv, clone, Manager, launch script. Act II is the animated map that connects install → checkpoint → node chain → Queue Prompt. Paste the commands first; the scene shows where each step lives in the graph.

Setting Up ComfyUI 🛠️

You can find the full setup guide in the official repo: ComfyUI on GitHub.

Here’s the quick path to a working comfy command: ⚡

Set up a virtual environment and install ComfyUI:

python -m venv comfyui_env
source comfyui_env/bin/activate
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.3
git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
pip install -r requirements.txt
cd custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Manager.git

Create a run.sh script to update & run ComfyUI:

#!/bin/bash

cd /home/hoangyell/Projects/ComfyUI
git pull origin master
source ../comfyui_env/bin/activate
python main.py

Add an alias to your ~/.bashrc to run ComfyUI easily:
- alias comfy='/home/hoangyell/ComfyUI/user/run.sh'
- Then reload your shell: source ~/.bashrc
- ✅ Now you can launch ComfyUI with: comfy

Act II — the map. You’ve got the shell commands. The scene below is the same journey as a node graph: install, pick weights, optionally steal a Civitai Red workflow, wire the core pipeline, write a prompt, add LoRA, hit Queue Prompt. Eight beats—bookmark this when the canvas looks like spaghetti:

Why ComfyUI — modular beats dead A1111 forks
Local setup — venv → Manager → python main.py (what you just pasted)
Pick checkpoint — SDXL / Flux / SD 3.5, with a VRAM gate
Import workflow — drag a Civitai Red PNG onto the canvas
Node pipeline — Load Checkpoint → CLIP → KSampler → VAE → Save
Write prompt — CLIP Text Encode is your spellbook
Add LoRA — Flux-compatible style layers only
Queue Prompt — denoise, preview, export to outputs/

Beginner ComfyUI flow: install via venv and Manager, pick a checkpoint with VRAM-aware Flux paths, import Civitai Red PNG workflows, run Load Checkpoint through Save Image, write prompts, add Flux LoRAs, and queue generation.

When a beat ends and you want the written deep dive, jump here:

Scene beat	Section below
Pick checkpoint	Which Checkpoint + Hell Mode model links
Import workflow	Civitai Red drag-drop + Hell Mode gallery
Node pipeline	Terminology table (advanced labels)
Write prompt	Prompts: Your Magic Spellbook
Add LoRA	LoRA section
Queue Prompt	Hell Mode — first real generate

The spine to memorize is the Generate group: Load Checkpoint → CLIP Text Encode → KSampler → VAE Decode → Save Image. Every workflow you steal from Civitai Red is just a fancier wiring of that chain—the Import workflow beat is the cheat code.

Civitai Red vs civitai.com (2026 split)

In April 2026, Civitai split into two domains:

Domain	What it hosts
civitai.com	Safe-for-work (SFW) content only
civitai.red	Full catalog—including mature models, generators, and community features

If you are reading this post for uncensored Flux, spicy PNG workflows, and NSFW LoRAs, browse Civitai Red. The main .com site will not show the catalog you came here for. Drag-and-drop workflow theft works the same—just source the PNG from .red.

Which Checkpoint should you use? 🧩

Model	Parameter Size	Strengths / Use Cases
SD 1.5	~1 Billion	Fast performance on older hardware; supported by a vast ecosystem of LoRAs and ControlNets.
SDXL 1.0	~6 Billion	Excellent for high-resolution, general-purpose image generation with strong aesthetic quality and prompt adherence.
SD 3.5 Large	~8 Billion	Produces high-quality, detailed images with superior prompt understanding and generation fidelity.
Flux.1-dev	~12 Billion	Advanced image generation model with larger context understanding and state-of-the-art output quality.

If you want the best sexy and realistic pics, grab Flux.1-dev. Yeah, they say it’s censored — but with the right tricks, you can get as filthy as your heart desires.

SD 3.5 Large is also amazing for realistic, spicy generations.

The Pick checkpoint beat already showed the VRAM fork—full flux1-dev above 16 GB, flux1-dev-fp8 if you’re lighter on VRAM. The table is the cheat sheet for which family to grab:

Honestly, don’t stress too much about picking the “right” model. Just scroll through Civitai Red and look for images that make you go “Woohoo!” Then do two things:

Download the model and related files linked to that image.
Many Civitai Red PNGs include node graphs — Just drag, drop, and boom — you’ve stolen someone’s whole workflow like a prompt pirate.

For me? Flux is hands-down the best. 🔥

Pro tip: 🕵️‍♂️ Got a model file (like .safetensors) but no clue what it is? Run this:

sha256sum your_file.safetensors

Copy that long-ass hash and Google it or drop it in Civitai Red search. You’ll find the model’s page and all the dirty details.

Prompts: Your Magic Spellbook ✨

In the scene, the Write prompt beat glowed CLIP Text Encode pink—that node is where your words become conditioning. Your prompt’s the magic wand; wave it right, and ComfyUI starts twerking. Here’s basic spellcasting:

ultra high resolution, 8k, amazing, realistic photo, ultra-detailed, beautiful face, white skin, long legs, photorealistic full body young cute Korean girl with a perfect body. She’s smiling. Sunlight hits her face from the left, making those shadows pop. Ultra detailed, sharp focus, pro photography.

Once you’ve got the basics, it’s time to flex like a true AI art thug: 💪

Those hotkeys matter on the canvas you saw in Queue Prompt—pan, zoom, duplicate nodes, save workflows:

Hotkeys for doing things faster: ⌨️⚡

Action	Shortcut
Pan Canvas	Middle-mouse drag
Zoom	Scroll wheel
Add Node	Double-click canvas
Duplicate	Ctrl + D
Delete	Delete key
Save Workflow	Canvas menu → Save
A Stranger Walked In	Alt + F4

Where to Go Next 🧭

The Node pipeline beat only lit up five nodes. Real workflows sprawl—this table names the rest when you’re ready to read the graph like a native:

ComfyUI Basic Terminology Table 📚

Term	Meaning
Checkpoint	Main model file (UNet + CLIP + maybe VAE).
UNet	“Draws” the image from noise during generation.
VAE	Compresses and decompresses images to/from latent space.
CLIP	Turns your text prompt into usable data.
Prompt	The text that tells the model what to generate.
CFG (Guidance)	Higher = follows prompt closely, lower = more freedom.
Steps	Number of image refinement rounds — more = better (usually).
Sampler	The method used to denoise (refine) the image.
Scheduler	Controls how noise is reduced — affects style and quality.
Seed	A number to repeat results — same seed = same image.
Latent	A compressed, abstract form of the image.
Tensor	Data structure for storing things like images and noise.
Conditioning	Tells model what style/content to focus on.
Concat	Joins multiple prompts end-to-end.
Combine	Blends prompts by averaging separate base noises.
Average	Averages prompt embeddings before denoising.
Loaders	Nodes that load checkpoints, VAEs, images, etc.
Latent ↔ Image	Convert between image and compressed form.
Preview	Shows you what the image looks like.
Save Image	Exports the final image to disk.
Node	A block that does one task (load, prompt, sample, etc).
Workflow	A full setup of connected nodes to generate an image.
Model Loader	Loads the checkpoint (.ckpt or .safetensors).
VAE Loader	Loads external VAE if not included in checkpoint.
CLIP Text Encode	Converts prompt to an embedding.
KSampler	Main node for sampling — uses UNet + prompt + noise.
Noise	Random input the model turns into an image.
Image Size	Resolution of the generated image — must be divisible by 8.

That’s a hell of a lot of info — but don’t worry, it’s for the advanced stuff. You don’t need to know any of that to have fun. You can always come back to it later.

The Game on beat was the trailer. Hell Mode is the first real generate—real Flux weights, real VRAM pain, real output.

😈 The Hell Mode 💭🔥

Once you’re ready to run ComfyUI for the first time with the basic setup, you’ll probably have some default models to play with. If not, no worries — grab them from the links below (but hey, I’m not responsible if you catch a virus 👀).

Then you can start generating some spicy angles like these: Just download the images and drag-and-drop them into ComfyUI — you’ll instantly get a full workflow, just like mine. That is the Import workflow beat in action. Grab those PNGs from Civitai Red—the SFW-only main site will not carry this catalog.

Looks amazing, right?

Nope, Too laggy? — your PC probably isn’t strong enough to run it with flux1-dev.safetensors. Told you already: if your GPU has less than 16GB of VRAM, you’re outta luck.

That is the Pick checkpoint beat sending you down the flux1-dev-fp8 branch—not a surprise, just physics.

But hey, don’t cry just yet — there’s still hope. 💧 Try the lighter version:

👉 ComfyUI/models/unet/flux1-dev-fp8.safetensors

Now, head back into ComfyUI, find the Load Diffusion Model node, load the file you just downloaded, and try again. 🔄 This one should run just fine — and boom, you’re finally generating your own sexy images, right? 😏🔥

Alright, I’m out. Or am I? 👀

Oh wait, not meet your expectations? You tried a naughty prompt to describe your dirty fantasy but the response is not good, your model always puts on a dress instead of taking it off like you wanted? 😂🤦‍♂️

That’s because Black Forest Labs wants to keep naughty folks like you out. 😏🚫 But lucky for us, the internet’s full of degenerates — and they made a new model just for you: 😈

👉 Flux Uncensored on Civitai Red

Pick a checkpoint from the search results—filenames vary by uploader (fluxUncensoredFemale_v10.safetensors is a common one). Download, load it in ComfyUI, hit generate, and… let your wildest fantasies run free. (Trust me, I’ve got some fantasies even dirtier than yours—but I can’t share them here! 😉)

💥 Welcome to hell. 🔥 Now you can create whatever filthy, wild imagery your wonderfully twisted mind can dream up.

LoRA 🎨

Hey, still around? Not quite satisfied because your images aren’t as stunning as you imagined? The Add LoRA beat showed where style layers plug in—between Load Checkpoint and the sampler. Wanna add some extra spice to your renders? 🎨

Looking for celebrity faces or signature styles? Say no more — LoRA is here to elevate your generations.

⚠️Note: If you’re using Flux, make sure to use Flux-compatible LoRAs only!

Here are some of my favorite LoRA picks you can try out: 🌟

Base LoRA – Shakker

Nanc / The Queen of Us – Shakker

Here you go—bet you’ll use this LoRA to swap in the face of your naughtiest characters. 👙

🧪 Pro Tip: This one shines when combined with:

Style Layer 1

Style Layer 2

Eunb on Stage – Shakker

Download my image, drag and drop it into ComfyUI to load the workflow—then turn the stage into a wild strip club for the ladies to go crazy! 💃🕺

X1 – HuggingFace X2 – Asian Female [flux] LoRA

Too spicy to display here! 🔥🙈

🙋‍♂️ Quick FAQ

Q: Is this legal?
A: Depends on what you generate and where you live. Don’t be dumb.

Q: Can I use ComfyUI on Mac?
A: You can try. But real degenerates use Windows or Linux with beefy GPUs.

Q: Can I sell the images?
A: If you didn’t use copyrighted characters or celebs — probably. Just make sure you follow the TOS for any models or LoRAs you use. If you did… good luck.

Q: Will this make me go blind?
A: Only if you don’t blink while staring at your generations for 12 hours straight.

Local ComfyUI vs Cloud Tools

Use ComfyUI when you want uncensored local control, Civitai Red workflow theft, and GPU-heavy Flux on your own terms.

Use ChatGPT / Midjourney / etc. when you want zero setup and you are fine with filters, quotas, and someone else’s rules.

This post assumes you already got blocked somewhere fun and decided to run the graph yourself. If the eight-beat scene, the spine, and the four-step exercise all clicked—you are graduated. Everything else is Hell Mode spice and LoRA garnish. 🔥