New Model(s) Planned for Release

I have been working this month on a released model merge which was planned to create a better illustrative model which combines the base model blend with FusionxRealistic - hoping to come out with a model which was better than the "Anime" model released under the Fusion blend of models.

The results:

A model which favors a drawn style but retains realism. I've been calling this style, "realistic anime" though there may be a better word for it.
A model which favors more realisticness but retains a semirealisticness which I call a "3DCG" style.

This is not to say that either model is incapable of producing a realistic or anime/line style drawing; however when you strip it down to the basic prompts you quickly discover what model style it favors. From this I can say that both of these models favor a "3DCG" feel, but one is more anime, and the other more realistic.

The tests:

I test with the same prompts, same seeds, same other settings and do a side by side comparison of each. Based on these results, I have narrowed it down to 2 final versions. Initially I had 3 versions which I was contemplating releasing just one, but due to the results noted (one anime, one semirealistic), I have eliminated one of the options and now we have just these versions.

I will do a final model test which will compare the same prompt/seed with all my models, then release all the images in a single post, and you can judge for yourself.

Model Limitations

Some downsides - the anime version has these negative qualities: bad eyes (different colors)

bad fingers. The realistic version has better hands but still struggles with nails not being very accurate. So some hand/finger prompting may be necessary.

Sidenote about negative prompting for each new model:

I'll attempt to come up with as suggested standard negative prompt for each one to address this. I usually remove jewelry as a negative prompt, but if you're familiar with prompting - order matters and sometimes removing it will change everything else as an unintended consequence. So I'll leave this as that and let's continue reviewing the tests.

Test 1

Test 1 was my tree test which the middle column performed the best in terms of realism.

As you can see, the far version has a light in the mid right quadrant that appears to be a bright light - not realistic. Between the other two images, you can't see probably, but in a side by side comparison, I liked the middle better. But I did like the yellowish moon compared to the white moon, that part I think is more realistic.

Test 2

In test 2, we eliminated the first version, in the image test it is named "ACR4".

I had to compress it since the original version as a png was 39,000kb. Here is a 50% compressed webp file . The realistic version is in the center, and the anime blend is in the 3rd column. I haven't come up with a snazzy new name for it yet.

Based on column 1, with 2 of the images with bad anatomy, I am going to archive the model (since it was a resource in creating the other 2 models, but ultimately by itself is not good enough). So that leaves the other 2 models. The middle column has more realistic qualities, whereas the model on the far right more anime qualities in comparison.

Test 3

Neither model is strictly anime or strictly 3DCG. But stripping the prompts from details (previous test 2) and getting down to minimal prompt style to bring out the natural leanings of each model...we can see that one favors 3DCG and the other favors a more drawn style, which I'm calling "realistic anime" - though I'm sure there's probably a better word for it.

Results - no major anatomy concerns on a simple test. The only issue remains is the fingernails and jewelry may not be desired. The next step is to try to prompt out of the image leanings of each model to see how each performs in creating an image style which it doesn't naturally lean towards (prompting for a drawn style or a photorealistic image) and comparing the results.

Test 4

In test 4, I created 2 different prompts, based on each model, then used the 2 sets of prompts made for each model on the other model. The prompt style was for an anime_style or drawn_style.

Note 1: I did find that adding in a prompt for "white_background" did create a more drawn feel for the image as you can see in the following images:

Same model, same seed, same size, same everything EXCEPT: one prompted for "white_background" and the other did not.

{upperbody}::adult woman, age21, posed::{standing, relaxed} ;, wearing dress::{{conservative, dressed}, sfw}, {hands, side}, anime_style,

Negative prompt: watermark, {child, teen},

wearing:: jewelry, earrings_;

{hands::wrong_finger_count, ;, {{stiff, hands},painted_nails}},

2girls,

heterochromia,

hats, cosplay, cape,

Steps: 30, Sampler: DPM++ 2M, Schedule type: Karras Exponential, CFG scale: 7, Seed: 37489514882, Size: 640x960, Model hash: 989b125599, Model: ACR_fp16__ACR4_fxr3_0.25, Clip skip: 2, Hypertile U-Net: True, Version: v1.10.1, Hashes: {"model": "989b125599"}

{upperbody}::adult woman, age21, posed::{standing, relaxed} ;, wearing dress::{{conservative, dressed}, sfw}, {hands, side}, white_background, anime_style,

Negative prompt: watermark, {child, teen},

wearing:: jewelry, earrings_;

{hands::wrong_finger_count, ;, {{stiff, hands},painted_nails}},

2girls,

heterochromia,

hats, cosplay, cape,

Again, if I prompted for a white background and didn't want a white dress, then I have to add in more information like green_dress, then if I don't want the eyes or hair to be green I have to add that, like in this prompt:

{upperbody}::adult woman, age21, posed::{standing, relaxed} ;, wearing dress::{{conservative, dressed}, sfw}, green_dress_;, blonde_hair, red_eyes, {hands, side}, white_background, anime_style,

Negative prompt: watermark, {child, teen},

wearing:: jewelry, earrings_;

{hands::wrong_finger_count, ;, {{stiff, hands},painted_nails}},

2girls,

heterochromia,

hats, cosplay, cape,

Note 2: This prompt was created for a specific model so

And side by side comparison for each:

Test 5

In this test, the only difference is changing the style from anime/drawn to "photorealistic_style". I used the green dress, blonde hair, red_eyes prompt for consistency, set it to a random seed count and batch of 4 to see on a wider basis how each model performs. The first model, which leans towards realisticness, drew each image and I noticed that the eyes were changed from the red prompt to a brownish color - which follows with the realisticness because humans don't have red eyes naturally. The anime leaning model drew in a more realistic manner but not photo realistically. Neither model with this prompt change created images that were photorealistic. But each tried their best.

Test 6

Final Landscape Test

Cover Image

I had my choice - so this is the other picture. The suggested resolution was 850 X 400, and this is a 2x. And here was the prompt:

{a tree::oak, autumn_; on a hill, daffodils}::{sky::midnight,moon_;, misty fog}):r_;

Negative prompt: watermark, {human, person, man, woman, child}

Steps: 30, Sampler: DPM++ 2M, Schedule type: Karras Exponential, CFG scale: 7, Seed: 124784286716, Size: 850x400,

Conclusion

In conclusion - These are 2 similar but different style leanings for each model. They can both draw in either style - but one favors and outperforms the other in the same style (the anime leaning version performs better in the anime category, and the realistic one performs better for realistic images). Using the same prompts reveal that each favor a particular style even when a style isn't prompted for.

Final Thoughts

I still need a name for each model and produce showcase images.

Feel free to let me know what you think.

Upcoming Models

New Model(s) Planned for Release

The tests:

Model Limitations

Sidenote about negative prompting for each new model:

Test 1

Test 2

Test 3

Test 4

Test 5

Test 6

Cover Image

Conclusion

Final Thoughts

Comments