nodes/Lex-DRL-ComfyUI-BestResolution
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

README: better phrasing

Browse Source
This commit is contained in:
Lex Darlog (DRL)
2025-07-06 14:17:03 +05:00
parent 94f110c405
commit 0eefe12f0e
1 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+4 -4
View File
@@ -60,10 +60,10 @@ This parameter has the highest priority. A desired resolution is slightly tweake
So, this imposes a few restrictions on the initial gen if we want to avoid cropping/outpainting later:
- The initial image must be a multiple of `8`
- After 1.5x upscale it must also be a multiple of `8` - so a +0.5 of size must be divisible by 8, or the initial step must be divisible by `8*2`
- After that 1.5x upscale we get into a territory where the ONLY next upscale we can do is 2x - which at that point might have not enough resolution to safely do so without messing up some details and having to denoise with high value again, losing some aspects of the base image we liked. That's a bummer...
- HOWEVER, if our initial `step` had another factor integer built into it, we can do a fractional upscale again.
- That factor could be `2` again... but after some tests I discovered `3` is a better choice. So, our total step needs to be `8*2*3` - then the second upscale could be +1/3 or 1.3333x
- Assuming we do at least one more upscale of 2x somewhere down the road, we can use the fact that we did a 1.5x before - since together they become another multiple of 3, or, in other words, the upscaled image could be divided by 9 on each side.
- After the mentioned 1.5x upscale we get into a territory where the ONLY next upscale we can do is 2x - which at that point might have not enough resolution to safely do so without messing up some details and having to denoise with high value again, losing some aspects of the base image we liked. That's a bummer...
- HOWEVER, if our initial `step` had another integer multiple built into it, we could do a fractional upscale again.
- This multiple could be `2` again... but after some tests I discovered `3` is a better choice. So, our total step needs to be `8*2*3` - then the second upscale could be +1/3 or 1.3333x
- On the other hand, if the second upscale was not fractional but 2x, we can use the fact that we did a 1.5x before - since together they become another multiple of 3, or, in other words, the upscaled image could be divided by 9 on each side.
- This is not at all necessary for "regular"-sized HD images, but might become handy for later ultra-upscales (6k+) with USDU.
Combining all of the above, our lowest common denominator is `8*2*3`, and assuming we do 1.5x + 2x upscales, it will naturally become divisible by `9` down the road.