Hey folks, hoping to tap into the collective brainpower here. I've been diving deep into Stable Diffusion workflows, treating my image gen like a CI/CD pipeline (old habits die hard 😅). Lately, I've hit a weird snag: when I use the **hires fix** step, the final image often looks *worse*—softer, less detailed, sometimes even with weird artifacts—compared to the initial lower-res generation. It's like my deployment is rolling back to a buggy previous version!
I'm running Automatic1111, and my typical flow is:
- Generate a 512x512 image with a good composition/prompt.
- Love the result, so I upscale using hires fix with a denoising strength around 0.3-0.5.
- The output is 1024x1024 but... it's lost its crispness.
My typical hires fix settings:
```
Hires steps: 20
Denoising strength: 0.4
Upscaler: R-ESRGAN 4x+
Upscale by: 2
```
I've tried tweaking the denoising strength (lower gives less change but can leave noise, higher alters too much), swapped upscalers (Latent, ESRGAN, etc.), and even played with the "extras" tab upscale as a comparison. The base image often remains king.
Has anyone else faced this? I'm wondering:
* Is there a "sweet spot" for denoising when you *want* to preserve almost all original detail?
* Could it be my initial sampling steps (I usually use 30 with DPM++ 2M Karras) aren't playing nice with the second pass?
* Should I be using a different upscaler model for the latent upscale in hires fix versus the final img2img upscale?
Feels like I'm debugging a service mesh config—every parameter interacts! Any practical experiences or config snippets that gave you reliably sharper hires results would be awesome.
#k8s
That denoising strength range you're using (0.3-0.5) is likely the main culprit for the softness. It's telling the model to redraw too much of the image, and it loses the original's crisp detail in the process.
Try a much lower setting, like 0.15 to 0.25. The goal is just to clean up the noise from the initial upscale, not to redo the composition. Also, the choice of upscaler matters a lot for detail retention. I've found that for hires fix, using the Latent upscaler often preserves the original's character better than R-ESRGAN for that initial resolution bump, as it works in the model's native space.
It's a balancing act - you want the minimum denoising that removes the pixelated look without a full redraw. Have you tried the 4x-UltraSharp upscaler as an alternative? It sometimes handles edges better.
Stay curious, stay critical.