Help with PC shutting off during generation.

[Jt747]

Hello, so I am new to Stable Diffusion and ComfyUI. My PC isn't the best, but I think it should be good enough to run this locally. I had an issue where I was able to generate the first image just fine, but the few I tried after that, my PC just quickly shut down in the middle of it. From what I understand, some PSU's can cause issues if they are underpowered or don't handle the power spikes that AI causes, but I think mine has more than enough power and is a decent brand. Also, during generation, I monitored my GPU temps, and they hovered around 65C according to the GPU software. I also tried undervolting the GPU, but I did not actually watch tutorials online to find the best settings to underclock and undervolt the GPU, and just used the default settings.

So besides the PSU and the GPU, what other things should I be checking to see what's causing the issue and fix it?

Current PC specs are:

CPU: Ryzen 5 5500

GPU: Radeon RX 6600

RAM: 16GB DDR4

HD: 2TB SSD

PSU: Enermax Revolution D.F.2 850 (which I think means 850W of power) which I think is overkill of rmy system and would be more than enough

[NspxMiguel]

An 850 W unit is ample for a Ryzen 5 5500 and RX 6600, but its rated wattage does not rule out a bad cable, connector, PSU, or wall-power problem. An immediate power loss with no blue screen usually points to a hardware protection event rather than ComfyUI itself.

I would narrow it down in this order:

1. Return the GPU, CPU/PBO, and RAM/XMP settings to stock. A guessed undervolt adds another possible source of instability.
2. In Event Viewer, check Windows Logs → System immediately before the shutdown. Kernel-Power 41 only records that power was lost; look for an earlier WHEA-Logger, BugCheck, disk, or AMD display-driver event.
3. Log sensors to a file with HWiNFO while reproducing it. Check CPU package temperature and the GPU hotspot/junction temperature, not only the 65 °C edge temperature.
4. Run separate OCCT tests for CPU, GPU, and VRAM, then the combined Power test. If only the combined test shuts the machine off, PSU delivery/cabling becomes the leading suspect. A failure in one isolated test points more directly to that component or its driver.
5. Reseat the GPU power connector and use a direct PCIe cable from the PSU. Also test a different wall outlet or power strip if possible. Run MemTest86 if the power tests do not reproduce it.

Also note whether it powers off completely, reboots instantly, or shows a blue screen. Those three outcomes lead to different next steps. If you can share the last relevant Event Viewer entries and GPU hotspot maximum, it should be possible to narrow this down further.

[Jt747]

Thank you for the help. I will try those things and report back. Also, from what I saw a few people mention SD isn't the best moedel for editing photos. Is that true? Is it best used for image generation instead? And lastly, will the results be better than the free models on Hugging Face like Klein? I got some decent results most of the time, but sometimes it would miss something I asked it to do, or it wouldn't get a certain aspect quite right.

[NspxMiguel]

Stable Diffusion is a model family, and the result depends heavily on which checkpoint and workflow you use. A general SD 1.5 checkpoint is mainly a generation model. Plain img2img can restyle a photo, but it may change details you wanted to preserve. For a local change, use a mask plus an inpainting model/workflow; ComfyUI's own guide shows that an inpainting-specific checkpoint gives more natural transitions than the general SD 1.5 checkpoint.

FLUX.2 Klein is not merely a hosted Hugging Face demo. The 4B weights are open under Apache 2.0, and the model supports both generation and image editing. It is a sensible editor to try, especially when prompt adherence or preserving reference content matters. It will not win every image: a masked inpainting workflow is often better when you need to change one small region while leaving every other pixel alone.

For your RX 6600, hardware is the practical constraint. ComfyUI's official Klein 4B guide lists about 8.4 GB VRAM for the distilled FP8 workflow, so an 8 GB card is right at or below that figure and may need model offloading, which will be slower. SD 1.5 inpainting is much lighter. I would start with the built-in inpainting template, then try the official FLUX.2 Klein 4B image-edit template after the shutdown issue is resolved.

“Better” also depends on the edit. Compare both with the same input and request, and judge three things separately: whether the requested change happened, whether untouched areas stayed intact, and whether artifacts appeared. If Klein misses one instruction, a tighter prompt or a mask may help more than switching to an older general SD checkpoint.

Official workflows:

• https://docs.comfy.org/tutorials/basic/inpaint
• https://docs.comfy.org/tutorials/flux/flux-2-klein

[Jt747]

Stable Diffusion is a model family, and the result depends heavily on which checkpoint and workflow you use. A general SD 1.5 checkpoint is mainly a generation model. Plain img2img can restyle a photo, but it may change details you wanted to preserve. For a local change, use a mask plus an inpainting model/workflow; ComfyUI's own guide shows that an inpainting-specific checkpoint gives more natural transitions than the general SD 1.5 checkpoint.

FLUX.2 Klein is not merely a hosted Hugging Face demo. The 4B weights are open under Apache 2.0, and the model supports both generation and image editing. It is a sensible editor to try, especially when prompt adherence or preserving reference content matters. It will not win every image: a masked inpainting workflow is often better when you need to change one small region while leaving every other pixel alone.

For your RX 6600, hardware is the practical constraint. ComfyUI's official Klein 4B guide lists about 8.4 GB VRAM for the distilled FP8 workflow, so an 8 GB card is right at or below that figure and may need model offloading, which will be slower. SD 1.5 inpainting is much lighter. I would start with the built-in inpainting template, then try the official FLUX.2 Klein 4B image-edit template after the shutdown issue is resolved.

“Better” also depends on the edit. Compare both with the same input and request, and judge three things separately: whether the requested change happened, whether untouched areas stayed intact, and whether artifacts appeared. If Klein misses one instruction, a tighter prompt or a mask may help more than switching to an older general SD checkpoint.

Official workflows:

* https://docs.comfy.org/tutorials/basic/inpaint

* https://docs.comfy.org/tutorials/flux/flux-2-klein

Hey, thanks again for the help. However, I have a few questions about doing the OCCT tests. So I ran a test of my CPU for about a minute and a half, and it showed no errors, but my GPU software showed the CPU temp climb to over 91c in that time. Also, when I go to VRAM page, it asks me to enter a number. Do I put the max amount of VRAM I have? And lastly, I don't see a page for testing the GPU. I just see VRAM and 3D Adaptive.

Any advice?

Thanks

Edit: I should also mention that the shutdown occurred without any errors or blue screens like it was just very quick and sudden like the PC was trying to protect itself. Also, you say my Hardware is right at the minimum limits and too offload some is that correct? If I do this then with the offloading, though, I technically should be able to run it locally? Also when offloading does that mean there os a limit on how much you can generate in a certain period? Oh and you ,emtioned in a previous reply about the quality of the output when image editing. It sounded like you were saying that these models should me able to do exactly what I want but that I have to get crafty with my prompts and be specific. Is that correct?

[NspxMiguel]

Stop the stress tests for now. AMD lists the Ryzen 5 5500's maximum operating temperature as 90 °C, so a reading above 91 °C after only about 90 seconds means the cooling needs attention before a longer CPU or combined Power test. It does not prove that the CPU caused the original shutdown, but it is already a problem worth fixing.

Check these first with the system powered off:

• Make sure the CPU fan is spinning and connected to CPU_FAN.
• Confirm the cooler is firmly and evenly mounted.
• If it was installed recently, check for protective film and enough thermal paste.
• Clean dust from the heatsink and verify that case fans have a clear intake and exhaust path.
• Return PBO/overclock/undervolt settings to stock.
• Use HWiNFO's CPU (Tctl/Tdie) reading to confirm the temperature; motherboard or GPU utilities can sometimes label sensors poorly.

In current OCCT, 3D Adaptive is the GPU core test. The separate VRAM test checks the memory chips. Do not allocate 100% of VRAM because Windows and the display need some. If OCCT asks for a percentage, start around 80%; if it asks for MiB, leave roughly 1–2 GB unused. OCCT itself notes that it cannot test memory reserved by the OS.

Once the CPU cooling is fixed, test in this order while watching temperatures:

1. CPU by itself.
2. VRAM at about 80%.
3. 3D Adaptive for the GPU.
4. The combined Power test last.

Stop immediately if the CPU reaches 90 °C again or the machine shuts off. A sudden black-screen power loss with no blue screen is consistent with hardware protection, so I would not run the combined test while the CPU already exceeds its thermal limit.

Offloading means ComfyUI moves some model weights from VRAM to normal system RAM and transfers them back when needed. There is no generation quota or time-period limit; generations are simply slower, and you need enough system RAM. ComfyUI also offers --lowvram and automatic/async offloading. This can make a model fit in less VRAM, but it cannot make an electrically or thermally unstable machine safe.

And no image model can guarantee every requested edit. Better prompting helps, but for exact local edits the larger improvement usually comes from masking only the area to change, using an editing/inpainting workflow, and keeping denoise strength low enough to preserve the rest of the photo.

References:

• AMD Ryzen 5 5500 specifications: https://www.amd.com/en/support/downloads/drivers.html/processors/ryzen/ryzen-5000-series/amd-ryzen-5-5500.html
• OCCT test descriptions: https://www.ocbase.com/personal-features
• ComfyUI low-VRAM options: https://docs.comfy.org/troubleshooting/overview