Subsequent image degradation after loading LoRAs (requires checkpoint reload)

[marika1978]

Hello, first of all, I would like to compliment Vladmandic. I know compliments are easy to give and might not mean much, but I really wanted to express my appreciation. Secondly, I would like to report an issue that, from what I've seen while browsing the discussions, already arose around 2023 (if I'm not mistaken). Practically speaking, for the past few days—and seemingly without any changes on my end, such as software updates, GPU drivers, etc.—whenever I load any type of LoRA, the very first generated images are perfect, but the subsequent ones degrade more and more. I must reiterate that this is a very recent issue and, as far as I know, is not linked to any hardware or software upgrades of mine. Is there a solution to this besides having to reload the checkpoint every single time? Thank you for your time.
P.S. I am not a native English speaker, so I am using an AI assistant to translate this. Please excuse any translation errors.

[vladmandic]

thanks!

typical workaround for your issue would be to disable lora fusing - it costs more memory, but avoids degradation.
its in settings -> networks -> lora native fuse & lora diffusers fuse

but in general, this is an issue, not a discussion item.
and really needs full log to say anything more on why this is happening.

[marika1978]

Key Findings from the sdnext.log

1. Hardware and Software Environment

• GPU: AMD Radeon RX 9060 XT (16 GB VRAM)
• Platform: Windows 10, utilizing the ROCm (HIP 7.2.0) backend instead of
  DirectML.
• PyTorch: Very recent developmental/nightly version
  (2.11.0+rocm7.13.0a20260416).
• Significance: Running ROCm on Windows for AMD GPUs is an actively
  developing, experimental setup. Memory allocation and PyTorch's mathematical
  operations on this configuration can behave differently compared to standard
  NVIDIA/CUDA setups. These environment details are crucial for reproducing
  the bug.

2. Confirmation of the LoRA Fusing Behavior

• Log entry: 2026-05-27 16:37:26,296 | ... | Network load: type=LoRA
  networks=[...] method=native mode=fuse ... reason=content changed
• Significance: The log confirms that LoRAs are being loaded in mode=fuse
  (native fusing), meaning the LoRA weights are merged directly into the base
  model's UNet. On this experimental nightly build of PyTorch for ROCm, the
  "unfusing" operation (the mathematical removal of the LoRA weights after
  generation) is likely failing silently or leaving residual values.
  Consequently, with each subsequent generation, the LoRA weights accumulate
  over themselves, causing the progressive quality degradation.

3. Garbage Collection (GC) and Memory Offloading

• Log entry: GC: current={'gpu': 6.04, 'ram': 11.1 ...} gc={'gpu': 2.11, ...}
  fn=restore:process_images_inner why=final
• Significance: These lines illustrate how SD.Next attempts to free and clean
  GPU memory after each generation cycle. Monitoring these logs helps diagnose
  whether the AMD VRAM is being cleared effectively or if a persistent
  state/memory leak is occurring between runs.

I managed to extract this, though I'm not sure how revealing it is. Thanks again.

[vladmandic]

Significance: The log confirms that LoRAs are being loaded in mode=fuse (native fusing), meaning the LoRA weights are merged directly into the base model's UNet. On this experimental nightly build of PyTorch for ROCm, the "unfusing" operation (the mathematical removal of the LoRA weights after generation) is likely failing silently or leaving residual values. Consequently, with each subsequent generation, the LoRA weights accumulate over themselves, causing the progressive quality degradation.

yup, this is exactly it.