ComfyUI 27 Incredible HiRes with Kohya Deep Shrink (free workflows), Stable Diffusion
Вставка
- Опубліковано 6 вер 2024
- The Kohya Deep Shrink node, that currently resides in the ComfyUI 'for testing' folder, makes it possible to generate images more than twice the default size, without upscaling. The images show more detail and look more photorealistic, even at default size, and the bonus is a speed increase of 15-20%
Link to workflows:
drive.google.c...
Rudy your dedication to understanding the details of Stable Diffusion and Comfy UI is truly marvelous. Always look forward to your videos.
Much appreciated!
The problems with persons and Kohya Deep Shrink can be remedied somewhat by setting end_percent in the Kohya Deep Shrink node to 0.6
Thanks, I'll give that a try. Unfortunately there seldom is thorough documentation ... it's all experimentation and tinkering with the parameters is time consuming.
Nice job man ! Very good explanation and instructions.
Just one thing, next time please could you open the images even in the brower with "open image" so we can have an idea about details because non zoomed + YT compression, i personnaly don't see any diff !
I subbed :)
For who wonder is also working with animatediff, the time gain is interesting alltrought the look change a lot if you want to use do not plan to add after you find the prompt and the seed.
Another usage is to add the Kohya Deep Shrink only for the upscaled pass of the video
Thank you Rudy! Well explained and the landscape images that cames out of my ComfyUi are truly astonishing!
Glad it was helpful!
Thanks for the video, it was really helpful. I am still struggling with a Python vEnv issue and I can't get ComfyUI to run but I'll eventually solve that.
Since I'm a longtime photographer, I was thinking about a few things that you might want to look for in faces, if I may intrude.
Kohya introduces some geometry shifts. You can see that in the kitchen very clearly -- the two ovens (or whatever they are) to the left have not been generated on a straight line. In that case I believe Kohya sorted it out but it's a recurring problem with perspective and AI, it's very often misaligned. The same applies to the street/ramp that appear in place of the original water on your last example. It's harder to see as Lady in Red is blocking the end of the image but I think it's not 100% ok. I would like to keep the water so typically I would go to Photoshop and fix that...
The portraits are AMAZING!, I will write about that after I manage to install and test things here, but thanks in advance!! You are right in saying that faces look much better, but do pay attention to asymmetrical deformations introduced in the right eye vs left eye tilt angle. If you know how to use Photoshop, that can be corrected, but it wasn't present in the original images you showed, so Kohya is creating this (as well as other, more pleasant) deformations.
Another tip that might be more helpful - sorry, I'm picky!!, that's why I mostly generate "weird" stuff on AI!! -- ... What you call "pop" when running self-awareness seems to be mostly what Adobe does with it's really badly designed "contrast" filter. It increases BORDER contrast. You can see that quite clearly on the clouds in some of your first images -- there is a rather strong (and, to me, very annoying) black line around the clouds against the sun. It's AMAZING that Kohya can render all those details much better without resorting to edge enhancement.
Again, I don't mean to be obnoxious and I'm writing this here because I know you pay attention to details! Many thanks for the video and I look forward to getting some of my photo images as sources for the whole workflow!
Thanks for your feedback and your valid observations. AI can generate wonderful artwork, but it still struggles with realistic photography, although things get better with almost every new release. Yes, the Kohya node can deform, faces can become stretched or skewed. Tuning the 'End-percent' parameter can help. And also with the Self-Attention Guidance node the motto is 'use with care'. Lol, as a photographer you're going to hate the new video that is due any day now ... about postprocessing and adding noise. :)
Great job, it's always a pleasure to watch your videos, many explainations and examples but always with simplicity
Keep it up !
Thank you very much.
Love your vids. Thank youn Rudi.
Much appreciated
The faces look awkward though with Kohya despite the details.
Indeed, more often than not close up portraits get somewhat deformed. I'd use Kohya for non human renders until maybe there's an upgrade.
@@rudyshobbychannel maybe it could be couple with a face detailer ?
Did you take English classes with Mark Rutte? 😁. BTW you keep talking about self awareness, it's Self Attention.
Great discovery!! Can this be used for img2img creative upscaling?
I haven't tried yet, but my bet is it might just work to sample a 2048 pixel image that was offered to the Ksampler (as latent)
Very interesting , could it be used with controlnet ?
I haven't tried yet.
amazing channel
Thanks.
great tutorial! while keeping this high resolution, how to integrate Img2Img? or Inpaint? Is working at this resolution possible for Img2Img? how to integrate it... thank you :)
For img2img just start with a large image, say 2048px, and offer that to the sampler as latent after a VAEencode. The Kohya node site between checkpoint loader and sampler. May want to play with the Koyha node 'end-percentage', people and faces may get less distorted if you'd increase it some.
i try this, thanks
Hello again, I would like to emphasize again your workflow which I find great.
But to manage my composition, I would like to integrate 1 or 2 controlNet based on a basic image representing my composition.
Make image2image to direct my composition but keeping your workflow and SDXL quality
I still have trouble understanding how all nodes can interact with each other, i m a newbee...
Thanks for your help !
@@koumac Videos 12 and 16 are about controlnet.
@@rudyshobbychannel top ! many thanks !!
Anyone tried to implement animdiff here? Please help me 😚
Wow this is amazing, I can even generate hires on my older laptop now!!
Hey, could you pleaeaeaease implement animdiff? I tried it but get blurry video's, don't know what I did wrong here. Thanks Rudy!
Thanks for your feedback. I'm not into video myself so there most probably will not be videos on animdiff from my side.