This LoRA can do a lot of good things such as detail enhancement, creative upscaling, photo restoration and quality improvements.
It's a Image-to-Image(I2I) edit LoRA, so it will require image input, but you can use with Text-to-Image as well.
I will be honest here; model itself is quite capable of improving the image, just by prompting it, without needing of this LoRA, yet this LoRA improves the quality and guidance of the model itself on these tasks because it's trained on a dataset that contains pretty good pairs.
If you are aiming for upscaling to higher pixels, please use this very simple workflow + add Load LoRA node between Load Diffusion Model and KSampler as usual and load this LoRA.
If you are not aiming for higher pixel upscaling, use it as usual without needing the workflow above.
Trigger word is: enhance this image, yet i strongly recommend putting your own words/sentences afterwards. Keyword alone might not give you the look you are after.
For example I'm using this one as a starter, but this prompt doesn't work well with the illustrations, sometimes realistic photos as well, so I'm deleting the "sharpen textures", so please find your own version of the prompt if you are not happy with the outcome:
Enhance this image to high resolution with rich fine details. Sharpen textures and surfaces, add natural microdetails, and improve clarity and definition across the image while preserving the original composition, lighting, atmosphere, and visual style. The input may be an illustration or a real photograph; preserve the original style.
Trained on AI Toolkit by Ostris
Description
FAQ
Comments (10)
Works like a charm! I use it at lower weight with my own upscaling prompt for that additional fine detail finishing touch and it really does make a difference. Thank you!
Mind sharing the upscale prompt? <3
unfortunately in my detailed Side by side testing this lora just degraded and smoothed out any details vs just using the base model without the lora: https://civitai.com/posts/31181669
Any increases in details in the showcase images is purely from the workflow and would be more detailed without the lora.
Can you please avoid leaving negative reviews based on your experiments and you are not sure about it, especially blaming me with using workflow tricks? At this point, you’re spre4ding information that I believe is incorrect, both here and on Reddit.
I’ve added more examples here, including some taken from your profile, along with my workflow snapshot.
This LoRA was not trained on high-quality image pairs. Its core purpose is to improve the detail and overall quality of lower-quality/less detailed images. If you input an image that is already high quality/2K, I can’t guarantee that the LoRA will provide additional improvement in every case.
Please take a look at these examples and let me know if you still feel there is no improvement: https://imgur.com/a/HJjX5XI
@reverentelusarca The lora adds no improvement on top the base model in my testing, turning it off and using the same workflow makes a better more detailed image, I believe I was quite methodical in my testing, using some detailed Qwen 2.1 images and some lower detailed SDXL images.
@reverentelusarca I will try with the bf16 Qwen 2.1 model tomorrow in case it is lora incompatibility with int8, but if it is that might be the death of Qwen 2.1 if loras do not work well with Quants.
@J1B Did you completely miss his point about starting with an already sharp and detailed image? Without testing, I would assume that a LoRA (that is suppose to correct abnormalities) would assume that your perfectly focused, highly detailed image needs to be softened as a corrective measure.
Perhaps you should try using this LoRA on an image that has smudged elements, poorly focused, and/or JPEG artifacts to see if this LoRA is, in fact, doing exactly what it says on the tin.
Besides, why on Earth are you trying to enhance detail on the images you presented?! They're already detailed to the point of appearing unrealistic.
@Unperson I think you missed the point that my sbs are outputs of smooth skinned sdxl images with the qwen 2.1 model workflow with and without the lora and not before and after. I will do some more tesring and add a 3 panel before/base/base +lora. But right now my testing shows that if you want the best image turn the lora off. Training a lora is hard, especially on a new model where the community hasn't found the best hyper parameters yet.
@J1B I completely understand your point and agreed upon hyperparameter tuning. I'm heavily occupied this couple of days and I would apricate further feedback and bf16 experiment you are having.
For the community; his point is the model itself is already capable of doing these low-to-high improvements itself without needing of the lora. Even, using lora sometimes reduces the detail compared to what model natively can do.
Based on my experiments before releasing the model, I also noticed model itself is sometimes beats the lora enabled, but that was rare on my cases and I indicated in the description as "model is already capable, this lora is just a helper-guider for better understanding what to improve", yet J1B's own experiments says native model is beating the lora all the time. If anyone else having this type of experience please do share so we can make this model's v2 better than current v1.
@reverentelusarca Yes I did some more testing today and can see that it adds extra freckles and skin wrinkles vs the base model, but also can give a slightly AI/CGI look to skin, this was a problem when applying loras to ZIT. possibly some of the training data is synthetic and that is throwing off skin?