Four New ControlNet LoRAs Just Dropped... Here's How to Get Them Running in ComfyUI
NEW ControlNET SDXL Loras - for ComfyUI
Stability AI just handed creators four new brushes. Canny. Colorize. Depth Map. Sketch. Each one a different way to bend light and shape into something that didn't exist five minutes ago. The install process? Three steps. Not three hundred. Three.
What You're Actually Getting
These aren't incremental upgrades. These are new creative instruments.
ControlNet LoRAs give you precise control over how Stable Diffusion generates images. Think of them as guardrails for your imagination... they let you say "follow THIS structure" while the AI handles the rest.
Here's what each one does:
- Canny ... edge detection. It reads the outlines of your source image and builds from there. - Colorize ... takes black and white photos or raw sketches and breathes color into them. Variations on demand. - Depth Map ... understands spatial relationships. Foreground, background, everything between. Your compositions stay grounded. - Sketch ... this one's wild. Combine two images with different styles and watch it generate something entirely new. Two visual ideas... one output.
That last one alone is worth the install.
Step One: Update Your ComfyUI
Before anything else... update. Compatibility matters.
Navigate into your ComfyUI folder. Find the update folder inside. Double-click `update_ComfyUI.bat`. Let it finish. Close the window.
That's it. No drama.
Step Two: Install the ComfyUI Manager
This is the tool that makes everything else easier. The ComfyUI Manager tracks your custom nodes, identifies what's missing from imported workflows, and installs dependencies for you. It's the stage crew making the magic happen quietly. ✨
Head to the GitHub page for ComfyUI-Manager. Scroll to the installation section and copy the `git clone` command.
Then navigate: `ComfyUI folder → ComfyUI → custom_nodes`. Click the address bar. Type `CMD`. Paste the command. Hit enter.
Let it download. Done.
Step Three: Download the Models
This is where you make a choice.
Hugging Face hosts two versions of each ControlNet LoRA:
| Version | Size | Speed | Precision | |---------|------|-------|-----------| | Rank 128 | Smaller | Faster | Good | | Rank 256 | ~2x larger | Slower | Better |
Know your hardware. If your machine runs hot on complex workflows, rank 128 will serve you well. If you've got GPU headroom and want maximum fidelity... rank 256.
Download whichever set fits your setup. Drop the `.safetensors` files into: `ComfyUI → models → controlnet`
While you're on Hugging Face, grab the workflow JSON files too. Canny, Depth Map, Recolor, Sketch Basic. These are ready-made templates from Stability AI... pre-built node configurations so you can start generating immediately instead of wiring everything from scratch.
The First Launch
Run ComfyUI with `run_nvidia_gpu.bat`. First launch will take a moment... the manager needs to set up.
Once loaded, click Load and open one of your downloaded JSON workflow files. You'll get a warning about missing nodes. Don't panic.
Click the Manager button. Select Install Missing Nodes. Find what's needed. Hit install. Restart ComfyUI.
BAM... your nodes load clean.
Running Your First Workflow
Open the depth map workflow. Select your ControlNet LoRA in the load node. Choose your checkpoint... SDXL base works, DreamShaper XL works, whatever you've got.
Load a source image. Set your positive and negative prompts. Adjust these parameters:
- Strength ... how aggressively the ControlNet influences generation - Start/End Percentage ... what portion of the diffusion steps use ControlNet guidance - Steps, CFG Scale, Sampler ... your standard generation controls
Hit Queue Prompt.
Watch it work.
The sample workflow skips the refiner and runs base model only... and the results are still remarkably solid. You can layer refinement in later as you get comfortable.
A Note on Compatibility
At time of writing, these ControlNet LoRAs work exclusively in ComfyUI. Automatic1111 hasn't caught up yet. If you've been on the fence about which interface to invest your time in... this is worth noticing. ComfyUI keeps getting the new toys first.
Why This Matters Beyond the Technical
Every new tool is really just a new question: What will you make with this?
The depth map understands space. The sketch LoRA blends visions. The colorize model resurrects old images. These aren't just features... they're invitations.
Finding that special place where work and play intertwine is magical for creating deep neural connections. That's true for AI models. It's true for us too. 🎨
The install takes maybe twenty minutes. The creative possibilities? Those don't have a ceiling.
New things are exciting because they hold potential. But how many new things have you let become old things without meaningful extraction? These four ControlNet LoRAs are sitting on Hugging Face right now... waiting. Twenty minutes of setup. A lifetime of creative exploration on the other side. Go build something. 🚀
--- Source: https://www.youtube.com/watch?v=sZrwDQ22Mr0
From TIG's Notebook
Thoughts that surfaced while watching this.
Progress, not perfection. Don't doubt yourself... doubt kills. When you pray for rain, you gotta deal with the mud too. — *The Equalizer series*— TIG's Notebook — On Failure & Perseverance
Having failures in life is important and having them early in life is a gift.— TIG's Notebook — On Failure & Perseverance
If you are able to emotionally heal and not allow it to turn into a bitterness, then it becomes a superpower. — *Chaplain TIG*— TIG's Notebook — On Self & Identity
Echoes
Wisdom from across the constellation that resonates with this article.
Audit which of the four prompting disciplines you currently practice and identify your weakest layer
Overlap identical parts and mirror symmetrical elements to maximize texel density without increasing texture size
It's not that cheap filters can't do the job, they can, but I can definitely notice a difference. I find it hard to explain, but every time I use them I found the image to be very pleasing.