[Bug] An unfinished line in the image generation (usually at the bottom of the image)

[alexdisablo]

Hi all!
I've been having a problem with image generation for some time using the stable SD client (I'll post details below). Long story short - in almost all image generations I have an "unfinished" thin line at the bottom of the image. Sometimes it also appears on the sides. The whole image looks good except for this thin line, but it's visible and looks bad. It can be fixed via inpainting, so it's not unreachable or something; it's something wrong with the generation.

My guess is that my laptop is switching the main GPU (RTX 4060) to a CPU inbuilt one from AMD. Otherwise, I have no idea why this is happening. Please help.

• I have searched in opened issues and didn't find a similar case;
• I have tried a fresh install (even tried reForged - the same problem);
• This problem did not exist ~3 months ago, I've checked a lot of my past generations - no sign of this;
• Tried changing models, prompts, Lora's, parameters of generations (like steps, image size), no effect;
• I see no error in the console.

Launch console log:
Full log + generation of 1 image:
Logs.txt

Launch log:

Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug  1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: f0.0.17v1.8.0rc-latest-278-gbfee03d8
Commit hash: bfee03d8d9415a925616f40ede030fe7a51cbcfd
CUDA 12.1
Installing forge_legacy_preprocessor requirement: changing opencv-python version from 4.10.0.84 to 4.8.0
Installing sd-forge-controlnet requirement: changing opencv-python version from 4.10.0.84 to 4.8.0
Launching Web UI with arguments: --cuda-stream --pin-shared-memory
Total VRAM 8188 MB, total RAM 15554 MB
Set vram state to: NORMAL_VRAM
Always pin shared GPU memory
Device: cuda:0 NVIDIA GeForce RTX 4060 Laptop GPU : native
Hint: your device supports --cuda-malloc for potential speed improvements.
VAE dtype: torch.bfloat16
CUDA Stream Activated:  True
Using pytorch cross attention
ControlNet preprocessor location: D:\SD_Main\webui\models\ControlNetPreprocessor
Tag Autocomplete: Could not locate model-keyword extension, Lora trigger word completion will be limited to those added through the extra networks menu.
[-] ADetailer initialized. version: 24.11.1, num models: 27
22:00:58 - ReActor - STATUS - Running v0.7.1-b2 on Device: CUDA
Loading weights [91b87bf726] from D:\SD_Main\webui\models\Stable-diffusion\[Pony] Perfect Pony XL.safetensors
model_type EPS
UNet ADM Dimension 2816
2025-01-02 22:01:04,821 - ControlNet - INFO - ControlNet UI callback registered.
[tag-editor] Custom taggers loaded: ['aesthetic shadow', 'cafeai aesthetic classifier', 'Improved Aesthetic Predictor', 'wd aesthetic classifier']
Running on local URL:  http://127.0.0.1:7860

Example of the issue:

[MisterChief95]

When you say you tried a fresh install, do you mean it has the same issue without any new extensions installed?

Also, have you tried it with the latest version of Forge? I can see you're using an old version from around June, probably for extension compatibility reasons, but if it happens in reForge and old Forge, but not new Forge, then that can help narrow it down.

[alexdisablo]

@MisterChief95 The same issue appears with and without extensions being turned on or installed.
I tried:

1. Turning extensions on and off on the old client;
2. Installing fresh version of ReForge client with no additional extensions (fresh vanilla install);
3. Installing fresh Forge client with last updates. It's more stable, but the problem still exists.

In all 3 cases the problem is still there

[MisterChief95]

All 3? That's not good. Can you share the PNG gen info for an image that exhibits this issue? Regarding your first comment, I can't imagine its swapping from your 4060 to the CPU for generations. Are you noticing abnormally long generation times at all?

[alexdisablo]

I can't say if it's abnormally slow, 1 image is generating ~ 1m 5-30 sec. Here are couple examples, description where to look + png info.
Although I must admit new Client is way more stable + without Lora this thing is almost invisible. I think it appears much more when I use Lora's (based on recent tests).

PNG info No Lora.txt
PNG info Lora.txt
PNG info no Lora (2).txt

[MisterChief95]

Does it show up as much if you disable Latent Modifier, FreeU and remove the embeds? Also, 1048x1392 is a bit of an abnormal resolution for an SDXL gen. Was this upscaled in any way, or is that the size you're directly generating at? Does it show at 896 x 1152 (same aspect ratio)?

[alexdisablo]

Yeah, though I may test a little more without embedding, that's a good point. Turning off FreeU, Modifier didn't affect this issue in any way. Tried with and without. Thx for helping me with this!

It's not an upscale, previosuly I was generating in a lower resolution, but then I saw the advice to make it higher for better overall quality. I can say that backgrounds are x2-x3 better now. And the problem existed with a lower resolution before, so this should not be the issue.

[MisterChief95]

Hm, I'm afraid I don't know what the cause is then. I tried reproducing the artifacting with as close a setup as I could get (same model, extensions enabled, embeds, resolution, etc.), but everything seems fine...

Some last ditch efforts:

1. Try installing Xformers instead of using PyTorch attention
2. Try using Emphasis without normalization ("No Norm"). Can be found in Settings > User Interface > Quicksettings List, add Emphasis, reload UI - will be on top bar)
   • Will affect generations, but No Norm is considered a bit better for SDXL anyway

[alexdisablo]

Okay, will do! Thank you for all your help @MisterChief95