My Workflow (Patreon)

Hello everyone! Today I want to talk a little about my workflow, especially when it comes to requests. Hopefully this helps some of you understand better what I can do, what works well, and where the limitations are.

An AI image model learns from millions of images during training, just like an artist learns by looking at other people's art.

It remembers patterns and styles, not the actual images, and uses what it learned to create new pictures.

The issue is that the AI can only work with what it learned during training.

If certain characters, clothes, or poses weren’t represented well enough in the dataset, it may simply not be able to reproduce them reliably.

That’s why I always ask people to check the character list here first: https://animadex.net/?mode=characters&page=9thats if the character was included in the dataset.

If the model doesn't know a specific character, there's another option called a LoRA. Think of it like a DLC for the base model.

It adds something new, like a character, art style, or outfit. But someone has to train that LoRA first.

You can find and download LoRAs here: https://civitai.red/search/models?baseModel=Anima&sortBy=models_v9&query=tsunade

Now to my actual workflow. It’s pretty simple.

You guys send me your requests, and I translate them into prompts for the AI model. Over time, I’ve learned which prompts are more likely to trigger certain poses, actions, camera angles, and so on.

Then I generate the first image.

Most of the time, the first result won’t look exactly like the request. So I keep adjusting the prompt, adding things, removing things, changing wording and run the model again.

It’s basically a lot of trial and error until I get something I’m happy with.

Once I like the image, I check it for mistakes. Things like bad hands, fingers, toes, limbs, or other obvious errors.

If the mistakes are small enough, I open the image in Photoshop and fix them manually.

If the errors are too severe to fix properly in Photoshop, I go back to the AI and keep generating until I get a cleaner result, or at least one that is easier for me to repair.

That’s basically my workflow for simple requests.

Now we get to the main problem: the more details you add to a request, the less likely it is that the final image will match exactly what you had in mind.

Every extra detail adds more complexity, and AI is still pretty bad at handling too many specific instructions at once.

Another thing to keep in mind: character-specific clothes, hairstyles, and similar details usually can’t just be transferred to another character with my workflow.

For example, I can’t simply put Tsunade’s iconic green coat on Hinata. There isn’t a prompt that perfectly describes that exact coat.

I can describe the general look of it, but it still won’t look exactly like Tsunade’s coat.

Again, the solution for this would usually be LoRAs:https://civitai.red/search/models?baseModel=Anima&sortBy=models_v9&query=tsunade

If it exists there, I can use it.

Hopefully you guys understand these limitations. Yes, there are other workflows, tools, and methods that can make certain things possible or get much closer to an exact request.

But that’s not the workflow I use, and it’s not something I plan to change.

If you can’t accept that, then please support creators who are able to replicate your request more exactly.

See ya!