This model is a multiple merge of various models carefully selected for their different abilities to produce realistic and detailed images. Capable of producing images in all styles, it remains primarily focused on the female anatomy and photorealism.
This model is the result of my long hours browsing new releases and testing dozens of models.
I can't give you the complete list of models or the recipe used to create it, but I would like to thank all the creators and contributors to this site who have provided me with so much inspiration and made it possible for me to deliver you this model.
I can only recommend that you use my workflow for ComfyUi which will guarantee you high quality images.
You can leave a review or tip me if you wish, it will be greatly appreciated.
Description
The adventure continues with this new version, it contains a touch of Illustrious which I think will bring better versatility and adherence to the prompt. Thank you all for posting your results, you're the best :)
FAQ
Comments (14)
The most genuine pubic hair. Hope there are more East Asian faces.
It's up to you to tell me. I also recommend modifying the detailers' prompt to get the best results on specific faces if you're using my WF.
As @graam558 said, you can modify the detailer prompt to generate a specific ethnicity better. With the upscale workflow, you can put (((East Asian woman))) in the detailer prompt to make it recognize it better. More brackets means more weight to the prompt, and SDXL really picks up on this.
V6 is a lot darker than the V5, and has higher contrast. Prompt adherence about the same, V5 has a touch better realism and a touch more diverse backgrounds. V6 is not a direct upgrade over V5, it's a sidegrade with dark bias and more contrast. V5 is the opposite: bright bias, less contrast, a touch more realism.
I agree completely.
@graam558
Can you show us the samples without
+ Lightning LORA
+ 2x Tile Controlnet
+ Two packs of eye, face, feet, hand, etc detailers
+ pack of hand, nipple LORAs
+ 4-5 KSamplers
+ Vintage Effect + Film Grain effect nodes
?
Your workflow uses at least 4 explicit samplers + 12 detailer's samplers. That is too much! + You use multiple LORAs.
How can we see what is the effect of your model?
Double pass is enough (T2I + I2I HiResFix) gives the correct impression.
Your workflow effectively takes an existing image, passes it through Tile Controlnet (which does an UNBLUR operation at its core). That shows nothing about the model itself.
The posted examples are all created using only my model. DMD2 LoRa is only used on the detailers for speed. The LoRas are only used on the hand and nipple detailers, and they aren't strictly necessary and don't affect the base image.
The images I generated and posted as examples follow a common prompt and workflow across different versions, allowing for comparison of the results.
There's no cheating with img2img, as you seem to suggest. Please refrain from trying to discredit my work, and as we say where I come from, if you don't like the soup, don't try to spoil it for others.
@graam558
>> Please refrain from trying to discredit my work
This is your words about cheating, and not mine.
What I trying to say is that it is really hard to understand how this model performs alone within simpler workflows.
https://civitai.com/images/117834517
I tried hard to decompose and understand the workflow of this example. In the beginning of the workflow you have input image (example.png). OK, there is a switch to T2I. What is the purpose of the controlnet (where the image does it takes from?).. It is sourced from sampler or from example?
There is so much noodles layered over your soup...
As we say where I come from, CAN WE TASTE YOUR SOUP without a need to dig throu all the other NOODLES?
@homoludens
Sorry, I took your comment a bit personally when I first woke up. I understand, my WF isn't the simplest, but I've always made sure it's organized and easy to read. The CNET file is used by Tiled Diffusion, which does a deep upscaling without losing detail or preserving the original image's character. For me, it's the best upscaling technique. You'll find all the details on the dedicated WF page, which I encourage you to check out. As for testing, I'll leave that to you; I don't use any other models besides my own. ;)
@graam558 I understand what Tile does. First impression was that this CN used on the input image. When I zoom out the WF I can't see what node what. Where are the samplers and what is the sequence. I am not arguing about this complex workflow for those guys who wants to push it to the limits.
But read my comment as a warning. Not so many civitans can into ComfyUI. They will use your model in A1111 WebUI and ForgeUI where they go with 2pass pipeline (T2I + HiRes Fix). They can bump into situation when amount of details and other features of their renders noticeably different from your samples. In that situation disappointment will probably cause the suspicions and questions.
This is why in the first post for the given version I prefer to use a basic 2pass workflow. Hope my point is more clear now.
@homoludens
Yes, I understand. It's true that this level of detail isn't achievable with Forge/A1111, which is why I switched to ComfyUI a long time ago, to get the most out of the models and explore the latest features. But I can no longer imagine generating images with a simple Hires Fix, it's become obsolete for me. I'm sorry for those who haven't yet taken the plunge. However, I provide all the necessary information on this page for those who might be interested :)
https://civitai.com/models/1837748/graams-wf-sdxl-hd-photos-advanced-worflow-for-hd-realistic-sfw-and-nsfw-photos
Hi graam558
Can you please help me out with Optimal settings for usage with Forge?, I see all your sample images are using Comfy and im not able to fetch the settings from them.
Thanks for your time
You can use the usual SDXL settings. Personally, I like to use the DPM++2m Karras sampler with 50 steps and a cfg around 4/5 for generating the base image, but many other settings are compatible. Since I no longer use an interface like Forge/A1111, I couldn't tell you which settings are possible and most suitable, feel free to explore ;)
@graam558 thanks, and finally any recommendation on the Upscaler type and Adetailer type please?
Thanks again!