Yes, I see your point. In practice, if you try to improve one aspect of the model, such as skin rendering, you inevitably lose out elsewhere. Achieving realistic texture essentially requires studio-shot datasets, which impacts the model's overall lighting. Conversely, opting for less detail means relying on outdoor photos without additional lighting, resulting in more authentic illumination. Making these choices inevitably entails compromises.
I will publish a V1 ConvRot version in the coming days.

Hi Zak, I’m really glad to see that the versions are sparking a debate. We all have different expectations, which is why our assessments differ. Personally, V3 is my favorite; it allows my LoRAs to fully express themselves—and does so in a truly masterful way compared to the other versions. Yet, V3 is the least appreciated and most criticized one.

I feel there is more to explore in the pictorial style than in realism, hence my interest.

Looking at my showcase, you can clearly see that the result is far removed from a "plastic" look. The issue stems from your prompt, which fails to leverage the model's capabilities, rather than the dataset, which is based on high-quality photos.

@Mikeflower Personally, I have nothing against plastic; it's just a texture like any other 😄

@Protonium I visited your profile to check the quality of your work; I always like to know who I’m talking to.

There are only 11 images on your profile, not a single one more, so I’ll refrain from commenting on or offering an opinion about them.

It is obvious that this profile was created solely to spew baseless criticism. Your opinion is being disregarded; it isn't worth dwelling on.

It will be done.

I’ve tested the "raw" model combined with "turbo"; so far, the results aren't conclusive. There are many parameters to fine-tune, and inference takes twice as long, so it’s going to be a long, drawn-out process.

I’ll most likely explore this hybrid approach further, even though the "turbo" issues are also inherent to the "raw" model. As for creativity, there are now tools available to compensate for that shortcoming.

Thanks for your feedback

Yes, I realize that using a double pass with the RES4LYF sampler (14 steps) makes things a bit slow.

On my setup, it takes about 1 minute 20 seconds—which is still acceptable—whereas a single pass with the standard sampler takes around 10 seconds.

There is another, less time-consuming dual-sampling approach that yields quite good textural results: the Two-Stage Sampler (check out the extension in the Manager).

@moutonrebelle I tested your prompt, and it is indeed very strange; there must be a pornography trigger word—the question is which one and why.

In any case, I’m going to rework V4; there’s plenty of room for improvement, and I’ll add this to the list of issues that need fixing.

That intrigued me, so I wanted to check. Sure enough, only V4 produces a pornographic result.

It seems that in this version, the interaction between a man and a woman in this position has been locked to a pornographic scene.

I should be able to pinpoint the source; I doubt it comes from a dataset—my suspicions lean towards one of the merged LoRAs.

For V4, I simply responded to a demand for optimized photorealism, which led me to remove the graphic and illustrative datasets.

I am certainly not abandoning the pictorial style—it’s my favorite—so I will revisit it in a future version dedicated exclusively to that look.

Clearly, it is quite difficult to optimize a single model for both types of rendering.

For now, V3 covers the pictorial and graphic styles well, while V4 handles photorealism, pending further improvements.

@Emenimuno I had already taken this approach with my Flux 1 models (Fluxmania), but I realized it was a mistake because it limited how users approached them.

When I work on realism, my testing focuses exclusively on that aspect; however, nothing stops a user from subsequently discovering a different visual style that appeals to them. So, the best approach is to leave usage open, while clearly specifying what the models were trained and tested on.

You haven't provided any details; I don't know which workflow or settings you're using, so how do you expect me to help?

What has been verified is that none of the four models generate noise when used correctly.

Throwing out comments like that serves no purpose.

@catmaxzj Which dual-sampling mode are you using? Clownshark samplers or standard samplers? What you're describing suggests a lack of denoising or poorly executed denoising. Areas to investigate include the number and distribution of steps between the two samplers, as well as the choice of sampler and scheduling. In any case, this issue is related to sampling, not the model itself.
I used dual-sampling with V4 (Clown Shark samplers) and achieved excellent results.
I also tested V4 with the double-stage sampler, which allows sampling to be distributed across two models and enables switching samplers mid-process. That works too.
Thanks for your feedback.

thanks for your feedback

yes

@williamyar9261 All my images contain their workflow; you just need to drag one of them onto the ComfyUI interface.

However, I have a pretty powerful setup (5090/32GB VRAM/64GB RAM), so my workflows are very heavy and VRAM-intensive; it’s possible they won’t work for you or is too slow.

@williamyar9261 I tweak my workflows daily, using single-pass, double-pass, or dual-stage setups, along with FaceDetailer, depending on my needs (if there’s no face, I disable or remove FD). If you check out my latest images, you’ll certainly see all these options in use. You can find examples of FD and dual-stage setups in the showcase for the recently published "KreaMotsar" LoRA.

@williamyar9261 I don't really agree with you; KR2 far outperforms ZIT. You probably have an issue with your settings or WF configuration.

I’ve also noticed this anomaly that pops up from time to time. It’s not exclusive to V4, I think it’s a general issue; I’ve seen it with other models too, so it’s likely an artifact linked to the native model itself.

This image was created using the workflow containing FaceDetailer:

https://civitai.red/images/136569652

Save it as a .png file and drag it onto your ComfyUI interface.