Testing Models for Realistic Male Figures

This short article takes a look at how well various models can pull off making realistic-looking males. Thanks to a spicy Reddit thread, I decided to get my hands dirty with some experiments – let's dive into the results and see which models are killing it and which ones... well, need some work.

Notes

• Important: This is my first post, so I'll refine my process for future tests. Currently, only the models and steps are provided for most of the images. The content of this article reflects my personal opinion. Feel free to express differing views and share your thoughts in the comments.

• Generation: These results were achieved using comfyui. I without inpainting, ControlNet, or IP-adapters, trying to keep it as simple as it can be. I'm using text_g and text_l.

• Prompts: I prompted until I got satifying results. For those that I already knew, it was faster, but for the ones I was not familiar with, it took a bit longer. Specially proteus which is one of it's own(style and prompting). Also, I used both text_g and text_l for the images EXCEPT for proteus that was giving awful results.

Models

1. Fluently_xl_v3.safetensors

2. truetolifesdxl_v12.safetensors

3. zavychromaxl_v60.safetensors

4. Proteus-RunDiffusion.safetensors

5. TempestV0.1-Artistic.safetensors

*I also tried newrealityxlAllInOne_30Experimental, but I wasn't getting good results with it.

Fluently_xl_v3

This freshly introduced model is proving to be quite effective for its inaugural release. While I initially faced challenges in obtaining satisfactory outcomes, and still have some reservations about its AI-like skin, overall, it's performing well.

• Steps: 30

• Resolution: 1024x1024

Truetolifesdxl_v12

This model has quickly become my preferred lightning-fast option. It combines the formidable capabilities of Juggernaut XL Lightning, RealVisXL v4 Lightning, and DreamShaper XL Lightning – three powerhouse models. The results are undeniably impressive, especially considering it only requires six steps of generation. The level of detail in the skin and overall anatomy is exceptionally high and promises further enhancement upon upscaling. Additionally, the model demonstrates remarkable versatility.

• Steps: 6

• Resolution: 1024x1024

Zavychromaxl_v60

This one caught me by surprise (shoutout to the person who brought it up in reddit post)! The skin texture, wrinkles, and overall facial anatomy are remarkably well-crafted, even when capturing subjects from a distance (a challenge for many models). Given its initial strong performance, I decided to push the model further by experimenting with less conventional compositions – and it handled them admirably. I didn't need to repeatedly prompt it to produce satisfactory results. With its high level of detail, I believe it's an excellent candidate for upscaling.

• Steps: 30

• Resolution: 1024x1024

***Proteus-RunDiffusion and TempestV0.1-Artistic***

These models, both crafted by DataVoid, appear to occupy a unique niche. They both necessitate additional steps, with Proteus's CFG being adjustable to a notably higher extent. Since its launch, I've been employing TempestV0.1 primarily for upscaling purposes. While the considerable number of steps may seem excessive, its performance is so exceptional that when paired with a sufficiently high denoise setting, it effectively rectifies facial and hand details while enhancing the overall image quality. Another noteworthy benefit is its utilization of higher resolutions, especially advantageous for tiling upscale methods, where using smaller tile sizes may lead to worsened image results and increased hallucinations. The major drawback, however, is the time it takes to complete the process. Nevertheless, the end results justify the investment, especially for those seeking superior upscaled image quality.

Proteus-RunDiffusion

Here we have it – the one that breaks the mold. The prompting process for this model is quite distinct and, from what I've read, takes some time to adapt to. However, once you achieve satisfactory results, it's undoubtedly worth the effort. The level of detail in the facial features is remarkably impressive, which could be attributed to its higher number of steps. The faces appear strikingly realistic. I'll certainly continue experimenting with this model.

• Steps: 50

• Resolution: 1024x1024

TempestV0.1-Artistic

This model was developed using extensive images (up to 4800x7200 pixels), resulting in intricate details and textures. As illustrated in the sample generation, the model performs exceptionally well in creating a wide range of male types.

*Avoid using the base model for img-to-img, as it tends to mute the color during subsequent passes

• Steps: 80

• Sampler: dpmpp_3m_sde

• Scheduler: exponential

• Resolution: 1536x1024

This was a revealing look at what these models can do, and the results have been surprising! Fluently_xl_v3, while still a bit rough around the edges, shows significant potential. Truetolifesdxl_v12 is a clear standout, delivering remarkable detail and versatility. Zavychromaxl_v60 consistently impressed with its realistic skin textures and ability to handle difficult compositions. The wildcard of the bunch is Proteus-RunDiffusion, which demands more work but produces incredibly lifelike faces. Finally, TempestV0.1-Artistic is a powerhouse in terms of its capability to generate details and its versatility..

This barely scratches the surface of generative models out there! If you think a specific model deserves its own spotlight, let me know in the comments. As always, remember, these are my opinions - I'd love to hear how your own experiments turn out!

Cheers.