8 steps vs 2.1 normal
Hi. I tested the 8-steps, and it's not good in all ControlNet modes. The only one that worked well was Canny; the others have artifacts or low quality. I will still test the 8-step Tile, but if the quality is really inferior, it might be worth considering training with a Tile without 8 steps.
HED 8 steps
HED Normal
Hello, the 8-steps distillation is the final step, so we originally had non-8-step tile models as well;
Looking at the results, it might be an issue with the strength setting. Could you share your workflow with me so I can take a look?
With this image for HED:
For 2.1 normal i used this:
prompt = "Cinematic film still, an ultra-realistic portrait of a melancholic Korean man in a dimly lit room. He is sitting at a dark wooden table, his hands wrapped around a green soju bottle. Rembrandt-style lighting, with a soft key light from the side sculpting his features and casting the other side in deep shadow (chiaroscuro). Sharp focus on his weary, expressive eyes. Shallow depth of field, with the dark background blurred out. Subtle film grain, art-house cinema aesthetic."
negative_prompt = "underexposed, crushed blacks, too dark, heavy shadows, makeup, smooth skin, plastic, wax, cartoon, illustration, distorted hands, bad anatomy, blur, haze, flat lighting"
target_height = 1024
target_width = 768
num_inference_steps = 25
controlnet_conditioning_scale = 0.7
guidance_scale = 2.5
For 8-steps
I tried the same config but with num_inference_steps = 8 with guidance_scale=2.5 or without guidance_scale=0
Hello, the 8-steps model was trained under cfg1.5 conditions. If your testing conditions are 25 steps + cfg 2.1, theoretically the results should already be quite good, but the speed is slower.
I just took a closer look, and I'm not entirely sure where exactly the 8-steps model falls short in details compared to the 25-steps model. To me, it actually looks fairly OK? However, you could also slightly reduce the control strength a bit more—you'll see that 0.6 and 0.65 in the images look more natural.
But I didn't run 25 steps in this 8-step model, I only ran 8 steps, and if I increase the control strength the result gets worse, if I decrease it the image looks more natural, but still.I think the image quality in the 25-step model is better (the one I called the normal HED). I'll do more tests tomorrow, thanks!
I also tested Tile 8 Steps and it's very good. See the image.
He only introduced an artifact next to the eye at that 1536 resolution; it looks good at 1280 resolution. If you could make the undistilled tile model available, I would appreciate it. Because I'm trying to build an image restoration pipeline for this model, and taking full advantage of the model's potential in this function would be great.
But I didn't run 25 steps in this 8-step model, I only ran 8 steps, and if I increase the control strength the result gets worse, if I decrease it the image looks more natural, but still.I think the image quality in the 25-step model is better (the one I called the normal HED). I'll do more tests tomorrow, thanks!
I apologize for the confusion. What I meant was if you're using the non-8steps version 2.1 model, then: If your testing conditions are 25 steps + cfg 2.5, theoretically the results should already be quite good, but the speed is slower.
The 2.1-8steps version works best at 8 steps.
What I'm actually more concerned about is which specific aspects you think the 8steps version is weaker than the non-8steps version, so I can identify and address the issues.
I also tested Tile 8 Steps and it's very good. See the image.
He only introduced an artifact next to the eye at that 1536 resolution; it looks good at 1280 resolution. If you could make the undistilled tile model available, I would appreciate it. Because I'm trying to build an image restoration pipeline for this model, and taking full advantage of the model's potential in this function would be great.
Thx.
Yes, the 25-step version is good. I personally think the skin texture is better, but it's a matter of taste. I also noticed that the 8-step version left men with more skin problems.
8-steps no CFG, 0.6 scale:
8-steps CFG: 1.5, 0.6 scale:
Tried 12 steps in 8-steps:
25-steps CFG: 2.5, 0.7 scale:
Comparison with zoom 133%
