UltraRes lets you render at higher resolutions than SDXL would normally allow. It supports txt2img as well as upscaling images in img2img to bigger image sizes than previously possible. Without the lora distortions occur at these resolutions.
The lora supports rendering up to 3 megapixel in txt2img and up to 7 megapixel in img2img. It was trained on Big Love XL2.5 with 130 images from the training sets of Subtle Styles and Sublte Poses. It is recommended to use it with Big Love XL2.5 and the DMD2 lora. For settings check the Big Love description.
Supported txt2img Sizes:
1536x2048, 1344x2016, 1280x1920, 1248x1824, 1536x1536 (and their landscape versions)
img2img Upscale Factor:
1.5 or 1.25
Recommended Lora Strength:
0.5 to 1.0: The higher the resolution, the higher the lora strength has to be to avoid distortions. Higher lora strength means less sharp details, so keep it as low as possible and as high as necessary.
ComfyUI Workflow
Download this image
Description
FAQ
Comments (20)
Hi! Great work! But I have saturation problems when i use mixed loras like "MultipleGirls", "Rayness2",and "DMD2"... So, how do you balance your loras and DMD2 weights? (i use biglove to, and my loras total weights are under 1.0) THX
Better reduce the weights further. A sum of 1.0 is usually too much. Big Love XL2 has less saturated colors compared to XL2.5. Might reduce it again in XL3. Another option is to prompt less saturation. Cloudy in the prompt often helps.
@SubtleShader Thx! I use xl2.5... I will try XL2 and the pony version too. Awating XL3 and keep up ur great job!
Again a great release! Will try it with my models...
Does this work with male focus or does it force women?
Yes, works with men too, although none in the training set. Just tested it.
That's great thanks, preview images looks really sharp and clean.
tested witth image to image, seems to work amazingly and fix a lot of stuff, less long torso syndrome, less double teeth in mouth etc, or I was just under a placebo effect, regardless good work :)
It fixes anatomy and adds more details, because it was trained on a much higher resolution with high resolution photos.
So i see you've recommended this be used with BigLove.
But I am using this with the Lustify Endgame model to interesting results, however with a lower 0.25 power than your biglove recommended 0.5-1.0.
Does this lora have to be used with hiresfix and upscaler to work correctly, or one can use it without?
Thanks.
Tried Lustify too, but got too grainy results. With 0.25 you get too many distortions at 1232x1816 and higher. Hiresfix produces lower quality in my experience. I use img2img. But you only need to upscale if 2-3 megapixel in txt2img (without hires fix) is not big enough. Then you can use img2img to get to 5-7 megapixel.
Hi all! i downloaded the workflow and all the models and loras but i have grainy white image (only noise) at 8 steps and 1 cfg with LCM exponential. any clue?
thank!
Sounds like you don't have or use the DMD2 lora
Hi SubtleShader, thank you for sharing the workflow. I'm having a problem because I'm only getting white noise when using it (I have DMD2 downloaded). Do you know what might be causing this?
Without DMD2 lora active you get very blurry results. Not white noise. Not sure what the problem is. Sorry.
Maybe Big Love was corrupted while downloading? Did you use Big Love before? If not, try downloading it again. One of the other loras (DMD2, UltraRes) might not have been downloaded correctly too.
Ok, seems a Comfy component was broken.
If I use this LoRA with another LoRA of a famous person, will your LoRA mess up the face and features, or does it just change textures and stuff like that?
It will likely change the face a bit just like any other lora would too (except for slider loras). But you can paste the original face back either in Photoshop or use a masking node in Comfy. You can also just upscale your image without this lora.
using the template provided every single image is out of focus... what is up?