Persistent faint border artifact on the right and bottom edges

[AndreyRGW]

Hi everyone,
I'm encountering a persistent artifact, and I've narrowed it down to the Hires. Fix process.
When I generate a low-resolution image, it looks perfectly fine. However, after the Hires. Fix pass, a faint grey border or padding appears on the right edge (and sometimes the bottom) of the final, upscaled image.

Here are a couple of examples where the artifact is visible on the final output:

• The artifact only appears when Hires. Fix is enabled. The initial low-res generation is clean.
• The problem is not model-specific. I've reproduced it with several different checkpoints, including ntrMIXIllustriousXL, cyberrealisticxl, and cyberillustrious_v50.
• Changing the VAE (or using the one baked into the model) does not solve the problem.

Has anyone else run into this specific issue with Hires. Fix? Are there any known workarounds or specific settings that might be causing this?

UPD:
The issue is caused by the final resolution from Hires. Fix not being a multiple of 8 (or 64). Stable Diffusion adds padding to non-compliant dimensions, which creates the faint border.

[w-e-w]

that's odd
for hires fix I'm pretty sure that we do a stretch resize so they shouldn't be any padding

can you share your exact generation parameters so I can recreate your situation and test it
also share the sysinfo found in setting tab > sysinfo se we ca see your settings

[AndreyRGW]

that's odd for hires fix I'm pretty sure that we do a stretch resize so they shouldn't be any padding

can you share your exact generation parameters so I can recreate your situation and test it also share the sysinfo found in setting tab > sysinfo se we ca see your settings

parameters
(medium shot:1.2) of a (BMW M5 F90), in (nighttime Tokyo:1.3) with (neon lights and bokeh:1.2) lora:dmd2_sdxl_4step_lora_fp16:1
Steps: 5, Sampler: LCM, Schedule type: Exponential, CFG scale: 1.5, Seed: 3935495600, Size: 1152x896, Model hash: 028106b7bc, Model: cyberillustrious_v50, FP8 weight: Enable for SDXL, Cache FP16 weight for LoRA: False, VAE hash: 235745af8d, VAE: sdxl_vae_fp16_fix.safetensors, Denoising strength: 0.4, Hypertile VAE: True, ToDo factor: 3, ToDo factor2: 1.0, ToDo disable after: 0.6, Hires upscale: 1.3, Hires upscaler: Latent, Lora hashes: "dmd2_sdxl_4step_lora_fp16: f032088e10cf", Discard penultimate sigma: True, Version: v1.10.1-73-g57e15ec9

sysinfo-2025-07-31-17-48.json