This model works as well in ComfyUI as in ForgeUI/Neo but is easier to use in ComfyUI. With ForgeUI/Neo it can make incredibly sharp outputs with HiResFix, but it's not as easy to use. Check my posts for comparisons.

I prefer ComfyUI because it can use the KL_Optimal scheduler. My tests show that KL_Optimal offers the best compositions. More facial variations as well and prettier women I believe. But Beta57, Exponential and Karas are often as good.

With ForgeUI/Neo I think you get less variety of faces than with ComfyUI, but using a character LoRA is an option. This model in Forge should mostly be used with Exponential, Karras or AYS, but many other schedulers can give good results too. Unfortunately KL_Optimal with LCM is a bad match for Forge because it forces automatic backend overrides behind the scenes, something ComfyUI avoids doing. ComfyUI is mathematically superior according to an LLM.

* Drag my PNG images in ComfyUI to get the workflow I used. But maybe you have even better prompts with upscaler and refiner.

* I have baked-in a skin LoRA for extra skin texture. However, if for some prompts the skin texture is too strong, you can load a skin LoRA with a negative value to reduce it.

* Avoid inside the prompts the words: Veins, Veined, or Veiny, since the model, like most other models sometimes put too much of that on the men's forearms (only the forearms for some reason).

The default settings, in my opinion, should be:

*In comfyUI:

LCM of course, with preferably KL_Optimal, Beta57, Exponential or Karas.

Steps at 10, 11 or 12 (8 is usually not enough) --- CFG 1 or 1.1 --- Clip Skip 2 is best, 3,4,5 can be good, (1 doesn't work).

Important: Use of a refiner/upscaler in your workflow is recommended as it acts like a second pass. Set HiRes CFG at 1. Try 0.26 or less for denoise and 5 steps or less. But that may vary depending on the scheduler used. Excessive denoise and steps in upscaler KSampler, can cause body deformations, bad fingers or deeper ribs, etc...

*In ForgeUI/Neo:

Same as above with Exponential as first choice, but other schedulers can do well. For best results you should activate the HiresFix tab using only ESRGAN_4x with 4 or 5 Hires steps, and denoise strength of around 0.24. Make the upscale size to 2048 or less.

I think using a Refiner can cause more harm than good in ForgeUI.

My strategy in Forge is to generate lots of images from one prompt without HiresFix and Refiner. When I get a great looking image, I send it from PNG_INFO back to Txt2img and tick HiresFix with the settings mentionned above so that the exact same image will be redone, but upscalled as in a second pass.

My preferred resolution with this model is (1024 or 1088) x 1344 as base resolution before upscale. Or you may prefer the more conventional 832 x 1216 + a bit more upscaling but I think that could lead to less sharp results.

***Tip you may not know and find useful: when you get an image you like but a little something looks wrong. Drag that image in comfyUI. Set the "Control After Generate" to FIXED. Then change something. Like removing or adding a comma or word in the prompt. RUN and see what changed. Or change Steps +1 or -1, or Weight_Interpretetion, or CFG to 1.1, or the sampler, Or change the clip Skip from 2 to 3. There's a chance you can get the resulting image you hope for. Having "Control After Generate" to FIXED, allows you to try multiple changes in the settings and still get a result based on the original seed. I can confirm that did salvage many of my posted images.

Same thing can be tried in ForgeUI. You drag the image in the PNG Info and send to Text2img. Tick the EXTRA box close to the original seed. Then you can set the "Variation Strength" you want. For example, 0.01, or 0.05, etc...