
Overview
I’ve utilized many AI tools to create a visual novel using my old comic as a source. You can read more about it here. You can download and play the visual novel for free here:
Project: A Day Of Fire And Fish
Warning, this is an adult visual novel! Do not play it if you don’t want to see naughty bits!
This post is an archive from CivitAi created May 22, 2026. It’s a short overview of how I used Anima to achieve the images I used for this project.
The Process

The release of Anima was the trigger for me to attempt this project. Anima offers much more control with tags and natural language thanks to it’s ability to understand spatial coherence. The 16 channel VAE offers crystal clear generation results which is exactly what I needed for my project.
Following in the footsteps of my previous projects, I trained a lora using a custom generated reference sheet in order to capture the details of each character. I used the original reference sheet for the character Ember, but had to create a new one for the additional character I added named Carla. Carla was the girl that appeared in the original comic but had only a minor role.


[These images were generated using Anima Preview 3. Ember’s character sheets were generated using a Noob SDXL based model. In hindsight, I should have corrected all of the imperfection in Embers images before I retrained them. I now know that Anima retains much more detail than SDXL. This oversight resulted in me having to correct every inconsistent detail that appeared in Embers images.]
Just like with my second comic, I chopped the reference sheet into smaller pieces to create more images for the lora training. Unlike either of my previous attempts, I ran into a problem. If I simply trained the reference sheets data, the backgrounds for the characters came out way too simple, which ended up clashing with detailed background of the in game scenes.

[One of the backgrounds from the visual novel. Created with Anima Preview 3 and MeMaAni P3]
So I came up with an effective but time consuming solution. I combined the character sheet data with the MeMaAni’s data and trained all of that together. It increased the already long training time but it somewhat fixed the background problem. Basically, I couldn’t create something as detailed as the background above with Ember present in the image. What I could do was something like this:

[The backgrounds didn’t get much more detailed than this. Which in all honesty, was fine with me. It fit the 2d style of the characters. I wish to state again, how well Anima (P3) replicates the characters. This is like 95% accurate to the training data.]
Thinking about it now, if I were to attempt it again, I would probably just generate the characters and the backgrounds separately and then manually put them together and train on that data. I believe that could solve the problem in the future.
Tagging
I don’t think I’ve mentioned this in previous post but better late than never. When it comes to tagging for characters, you want to tag for everything BUT the features you want associated for that character. You want those features to be associated with a custom tag you will use to prompt for that character under any circumstance. This is how I set up the tags for Ember.
fishluvr_ember,1girl,hooded black cape,white tank top,brown gloves,black shorts,belt pouch,brown boots,black thighhighs,light frown,standing,looking at viewer,arms at sides,reference sheet,multiple views,hand on own hip,straight-on,from side,white backgound,simple background,
fishluvr_ember is the tag I use when I want to generate her. Notice how I don’t list skin tone, hair style and color nor her eye color. All of that gets associated with the custom tag.
Sadly, this method isn’t 100% effective since the dataset is still on the small size. So as a means of reinforcement, I like to list the traits of the character when I prompt scenes for this. For example, here is the prompt for the scene in the image posted above.
fishluvr_ember, close-up, (upper body:1.5), three quarter view, foreshortening, Cinematic Lighting, black hair, red hair, gradient hair, twintails, long hair, brown eyes, small breasts, dark-skinned female, hooded black cape, white tank top, brown gloves, black shorts, belt pouch, red hair tie, slouching, sitting, in wooden chair, round cushion chair back support, arms behind head, legs on table, v-shaped eyebrows, frown,
Since the model knows what tank top, belt, twintail and all those other features are, it can automatically pull that data from the lora when you prompt for it. Now, it’s also possible for me to prompt for something completely different and get a custom design that still looks like Ember, thanks to the way I tagged the lora. For example:

[Using Amina 1.0 with a lora trained on Preview 3]
fishluvr_ember, dark-skinned female, 1girl, (solo:2.0), cowboy shot, white background, hands on own hips, embarrassed, v-shaped eyebrows, blush, looking at viewer, high ponytail, black hair, red hair, gradient hair, eyewear on head, brown eyes, bare shoulders, purple floral print, grey dress, off-shoulder dress, short dress, sundress, wavy sidelocks, small breasts, contrapposto,
Conclusion
Anima has been a game changer for game making. Had I been using Anima 1.0, I believe I would have had a much easier time achieving some of the more difficult multi-character compositions. If you want to learn more about why and how I made this project, read more here. Or just play it for free here.
I plan to make more projects like this going forward. Use the website I linked to keep track of my progress. This will not effect my lora making. I will still continue to make loras and post them as normal. The only thing I’m considering is making backup location for all of the loras just in case something happens to this account. So look out for that in the future I suppose.

