JA|EN

About This Guide

This is a build log for getting a local anime-illustration setup running with ComfyUI + Illustrious (SDXL family) on a newly purchased NVIDIA GeForce RTX 5060 Ti.

There are two things we want to cover:

  1. The setup flow itself — from download to first launch
  2. A pitfall specific to the RTX 50-series (Blackwell generation)

The second point is the important one. Miss it, and you’ll very likely run into a situation where the app launches fine but crashes the moment you try to generate. Steering you clear of that is the main goal here.

Who this is for: anyone on a Windows PC with a recent RTX 50-series GPU who wants to start generating images locally with ComfyUI.

The Bottom Line: What to Check First on Blackwell

On the RTX 50-series (RTX 5060 Ti / 5070 / 5080 / 5090, i.e. the Blackwell generation), you can’t generate images unless PyTorch is a recent build.

  • Rough requirement: PyTorch 2.7 or later, with CUDA 12.8 (cu128) or later
  • If you’re stuck on something older like cu124, the app itself launches, but generation crashes with this:
CUDA error: no kernel image is available for execution on the device

This happens because the GPU is too new for the older PyTorch build to contain the compute kernels for it. Internally, Blackwell reports as a new architecture tier called sm_120, and only a build compiled with support for it can actually run the math. This has nothing to do with the model you’re using — it’s purely a GPU-side requirement.

Since some distribution packages bundle an older PyTorch, don’t relax just because “it downloaded and launched.” Always verify the PyTorch version and confirm the GPU actually works. That’s the core of this guide.

Worth noting: this same failure isn’t unique to ComfyUI. Users coming from Automatic1111, Forge, or InvokeAI hit the identical no kernel image error on Blackwell cards — the fix is the same cu128 upgrade in every case.

A quick note from our own run: at the time of writing, the latest package already shipped with a recent PyTorch (2.13 line / CUDA 13.x), so it worked out of the box with no upgrade needed. That said, this varies by timing and package — we may have just gotten lucky — so it’s safer to always run the verification step regardless.

What You’ll Need

  • OS: Windows (this guide was done on Windows 11)
  • GPU: RTX 5060 Ti (16GB VRAM). The same reasoning applies across the RTX 50-series
  • Storage: a drive with room to spare. SDXL models are 6–7GB each, so plan for a few dozen GB to be comfortable
  • Install location: any folder (we’ll refer to it as <ComfyUI-root> throughout)

Tools you’ll want:

  • The latest NVIDIA driver
  • 7-Zip (the package ships as .7z)
  • curl (the one built into Windows is fine; resume support is handy for large downloads)
  • (Optional) Git (used when installing extensions)

Steps

Step 0. Clear Out Old Leftovers for a Clean Base

If you’ve tried ComfyUI before and it didn’t go well, layering a new install on top of a broken environment tends to cause the same problems again. Check for the following first, and set them aside if they exist.

  • Candidate install folders (anything corresponding to <ComfyUI-root>) and leftovers from the desktop app
  • Data worth preserving: downloaded models, generated images (output), and saved workflows and settings (user)

Rather than deleting outright, rename the old folder to isolate it and prepare an empty folder for the fresh install.

Our machine was clean, so we could skip this. The more existing setup you have, the more carefully you should handle this step.

Step 1. Check the NVIDIA Driver

Run nvidia-smi and confirm two things:

  • Your GPU (e.g. RTX 5060 Ti) shows up correctly
  • The CUDA Version in the top right is 12.8 or higher (i.e. the driver is recent enough)

If it’s old or unrecognized, install the latest Studio or Game Ready driver from NVIDIA’s site before moving on. On Blackwell in particular, a recent driver is a prerequisite.

Step 2. Install 7-Zip

Since the package is a .7z, install 7-Zip to extract it. A package manager like winget handles this in one command.

Step 3. Download and Extract the ComfyUI Portable Build (NVIDIA)

  • Grab the ComfyUI portable build (for NVIDIA/Windows, .7z) from the official releases (about 2GB).
  • A resume-capable download is reassuring in case the connection drops mid-transfer.
  • Extracting the .7z gives you the following under <ComfyUI-root>:
    • run_nvidia_gpu.bat (launcher)
    • ComfyUI (the app itself)
    • python_embeded (bundled Python)
    • update (update scripts)

The level containing run_nvidia_gpu.bat is what we’ll call the “ComfyUI root” from here on.

Step 4. [Most Important] Verify Blackwell-Ready PyTorch, and Fix It If Needed

Skip this step and, on the RTX 50-series, you’ll very likely hit a generation-time error. Tedious as it is, do it.

(1) Check the current PyTorch

Using the bundled Python, print the PyTorch version, CUDA version, and GPU status. Four things to look at:

  • torch version (is it 2.7.0 or higher?)
  • cuda version (is it 12.8 or higher?)
  • Whether the GPU is available (available = True)
  • The GPU name (does your GPU show up without errors?)

(2) Actually run a computation on the GPU (this is the decisive part)

Run one small matrix computation on the GPU. If a number comes back, you’re good. If you see no kernel image is available or similar here, it’s a fail.

Even when the version strings look correct, some problems only surface once you actually compute. available = True only tells you the driver detected the GPU — it says nothing about whether the compiled kernels actually run on it. So don’t skip this real-computation test.

(3) If it fails, fix it

Reinstall PyTorch from a cu128-or-later build. In practice, this means reinstalling torch / torchvision / torchaudio for the bundled Python from the cu128 index. When it’s done, run (1) and (2) again to confirm a pass.

The exact command varies by environment, so the official PyTorch “get started” page is the reliable reference.

✅ Pass criteria: torch ≥ 2.7.0 / cuda ≥ 12.8 / GPU available / the real-computation test returns a number.

Step 5. Update ComfyUI (and Re-Verify Afterward)

Use the update script in the update folder to bring the app up to date.

One caveat: as a side effect of dependency updates, PyTorch occasionally rolls back to an older version. After updating, run the Step 4 verification again to make sure you’re still in a passing state.

Step 6. Install an Extension for Easier Model Management (ComfyUI-Manager)

ComfyUI-Manager is the go-to extension that makes managing and installing models, LoRAs, and custom nodes far easier. Install it into the custom_nodes folder (Git makes this simple). Recent packages sometimes include it already, so the rule of thumb is: install it if it isn’t there.

Step 7. Download the Base Model (an Illustrious Checkpoint)

  • Location: <ComfyUI-root>/ComfyUI/models/checkpoints/
  • We went with Illustrious XL v1.1 (SDXL family, about 6.5GB). You can download it directly from the distributor (such as the official OnomaAI repository on Hugging Face) with no login required.
  • While you’re at it, set up the LoRA folder .../models/loras/ too (you’ll use it later when adding style LoRAs).

Corruption check: if the downloaded file is unusually small (say, under 1GB), it’s very likely a failed download where the contents are actually an error page (HTML). Re-download using a resume-capable method.

Why v1.1: it has broad compatibility with existing Illustrious/SDXL LoRA assets, so it’s easy to reuse what you already have. If you’re after a particular look, you can swap in a popular fine-tune (from the Illustrious family) later.

Step 8. The Default Workflow Is Fine

Illustrious runs on a standard SDXL pipeline, so ComfyUI’s default startup workflow works as-is. No special template needed.

  • Basic layout: Load Checkpoint → CLIP Text Encode (positive/negative) → KSampler → VAE Decode → Save Image
  • If you use LoRAs, building and saving a version with a single Load LoRA node inserted makes things easier down the line.
  • If you want clip skip 2 (a common convention for anime models), add a CLIP Set Last Layer node set to -2.

Step 9. Launch and Run a Test Generation

  1. Run run_nvidia_gpu.bat in the ComfyUI root.
  2. Launch succeeded when the console prints:
To see the GUI go to: http://127.0.0.1:8188
  1. Open http://127.0.0.1:8188 in a browser and select your model (Illustrious XL v1.1) in Load Checkpoint.
  2. Enter a prompt and press Queue.

The prompts we used for testing:

  • Positive prompt:
    masterpiece, best quality, amazing quality, very aesthetic, 1girl, long black hair, school uniform, cherry blossoms, soft lighting

  • Negative prompt:
    worst quality, low quality, lowres, bad anatomy, bad hands, jpeg artifacts, watermark, signature

  • Suggested settings: sampler euler_a / CFG 5–6 / steps ~28 / resolution 1024x1024

The first run takes a moment to load the model. If a .png lands in the output folder without errors, the setup is complete.

On our RTX 5060 Ti / 16GB VRAM setup, SDXL at 1024×1024 generated with comfortable headroom.

Troubleshooting (Blackwell-Specific)

  • CUDA error: no kernel image is available for execution on the device the instant you generate
    → PyTorch isn’t Blackwell-ready. The cu128 upgrade in Step 4 resolves it. This is by far the most common trap.
  • Torch not compiled with CUDA enabled / GPU unavailable (available = False)
    → A CPU-only PyTorch is installed. Reinstall per Step 4.
  • It suddenly stopped working after an update
    → PyTorch may have rolled back during the update. Re-verify Step 4 and re-apply cu128 if needed.
  • Your model doesn’t appear in the checkpoint list
    → Recheck the location (models/checkpoints/). A refresh or restart makes it show up.
  • The download keeps cutting off, or the file is suspiciously small
    → The contents are likely an error page. Re-download with resume support.
  • Out of VRAM (CUDA out of memory)
    → With 16GB, SDXL usually has room to spare. If it happens, lower the generation size or batch count, or launch in a low-VRAM mode.
  • A message says a specific .safetensors is missing
    → This isn’t a problem with the model itself — the loaded workflow is simply pointing at a model/LoRA name you don’t have locally. Reselect an existing model in that node, or remove the unneeded LoRA node, and it goes away.

Wrap-Up

  • The ComfyUI portable build is remarkably easy to get running: download, extract, launch.
  • But on the RTX 50-series (Blackwell), the PyTorch requirement (2.7+ / cu128+) is a hard prerequisite. This is the one thing you should verify with your own eyes.
  • With Illustrious XL v1.1 as the base and LoRAs added on top, you end up with a setup where generating is just a matter of changing the prompt in your browser.

The newer the GPU, the more that one extra step — checking whether the software has caught up — pays off. We hope this helps anyone else stuck on the RTX 50-series.

A note on licensing: Illustrious models and LoRAs each come with their own licenses. Follow the terms on each distribution page for whether commercial use and redistribution are permitted.