Fantasías Inanimadas

Tras probar de forma local sin éxito, pasé a entrenar Loras para Krea 2 en la nube consiguiendo bastante bien lo que buscaba.

Primero hice un par de loras de artistas que seguramente suba a CivitAI. John William Waterhouse y N.C. Wyeth. El resultado se veía claramente, pero los resultados tendían a ser demasiado oscuros y las caras no demasiado atractivas.

He conseguido mejores resultados incorporando además un Lora de imágenes sintéticas de Chroma que tiene unos colores más vivos y también he entrenado la cara de un personaje que ya tenía el set generado de una prueba que hice. Los resultados del entrenamiento de personaje son muy buenos, pero también se ha entrenado la ropa y el fondo (tiende a generar fondos verdes y el vestido a ser verde y rojo) No es muy difícil de controlar.

Otra cosa es que como no tiene negativos, tiende a firmar los cuadros debido al entrenamiento de los loras de artistas. Lo que he hecho es dejar a Krea 2 firmar los cuadros (Kreatwo)

Krea 2 queda pues incorporado para crear imágenes. Como lo único que estoy haciendo ahora son cartas custom de magic. Aquí van unas muestras.

La primera aprovecha la capacidad de Krea 2 para replicar famosos.

La segunda es una composición que creo que sería muy difícil que saliera bien en otros modelos.

Verricana, mi set custom de cartas de magic basado en Hollywood incluye también referencias a la época del Dustbowl y la gran depresión y a Las Vegas.

Trying to make Krea 2 work. Some notes:

Stylization is unreliable, faces are too doll like sometimes. The anatomy is perfect.

I need to train my Loras. In the meantime applying style with Chroma on the Krea 2 image is slow as hell but it's an improvement.

Más Krea2. El conocimiento de las caras de gente famosa es bastante bueno (varía claro según lo famoso) del individuo. No es que me interese demasiado esto. En estilos pasa más o menos lo mismo algunos los clava, como por ejemplo el manido Studio Ghibli, y otros le suenan. Este punto de partida creo que es bueno para que el resultado de los Loras con no demasiadas imágenes sea mejor que si los desconociera por completo.

Otro asunto es la censura aplicada al modelo que nunca es bueno. Parece que es una constante en los modelos de pesos abiertos que se haga un alineamiento de seguridad. Este parece ser más ligero que el que aniquiló a Stable Diffusion 3.

After a long hiatus for the completion of the visuals of my Visual Novel (here) I am back to writing.

I had several ideas but I have settled for what I think will be an easy task: writing a screenplay for "The Cats of Istanbul".

The Cats of Istanbul is a false film associated with AI imagen that appeared in the /tv/ 4chan board. A 1970s anime that had several sequels each more preposterous: the Cats of Istanbul in Russia, the Cats of Istanbul in America, the Cats of Istanbul travelling through time...

Even if my idea comes for the specific challenge of "What would have been that film about?" After having the premise, I've had the idea simmering for a while and I think that it is quite promising, to the point of leaving the meme aspects of the original idea aside. So without further ado:

This was made with Ideogram 4. It is quite good when you don't run into the censorship.

My visual novel is complete.

I realised that Chroma has been my goto model for more than a year, and after a string of new model releases, I just don't find it to be replaceable.

With all of its limitations compared to more modern models, it achieves stylization better than none and that is what matters most to me.

Believe me, I looked and tried to make my own loras for stylization elsewhere

Z-Image, Flux 2 have been a disappointment in that regard. They have their uses but I can't accomplish what I like. It's not only style, but these new models converge too much regardless of seed.

Maybe if I liked Anime I would have received Anima with open arms, but after briefly testing I didn't see the point for investing in a new model.

ChatGPT (in its own words the native Image capabilities of ChatGPT 5.x) offers even more capabilities regarding complexity and editing, but censorship and it gives a mostly safe and bland style. Getting it to edit an image while mostly respecting the style is the best thing it offers for me.

And what does the future hold?

Yesterday I decided to give Ideogram a go. Same prompt with Ideogram and ChatGPT

Disregarding the issue of composition with Ideogram (the image was turned to the side instead of using a panorama orientation as I intended), it's as accurate as the ChatGPT image, but instead of being a regular local model with better or worse NSFW capabilities it comes censored in the weighs.

The week went by very fast.

The visual novel implementation with Renpy is progressing fast: seven scenes to go.

I received the physical cards of my latest magic custom set, and I have decided to finish my Noire Fantasy Hollywood set, I have 50% of the cards already.

As customary the first Friday of the month I look to Imagen of previous years.

2025

I was already using the first ChatGPT model alongside Chroma. ChatGPT's previous imagen model had issues with colors and texture.

I was using Chroma and I am still using Chroma today. Will 2027 be different? I'll write my thoughts in the next post.

2024

SDXL

And SD3M

I was creating more complex scenes, but it was a flawed model.

2023

SD 1.5

A mix. Busy week.

Almost done with asset generation for the Visual Novel and I decided to try Stable Audio 3 now that I have more time in my hand; at the very last minute but another model contributed to this work.