V2
This is a complete revision of the original dataset, now composed entirely of custom-made material created specifically for this Lora, using real, live models (with very little synthetic material, less than 10%).
In this regard, the first thing to do is to give credit to those models who chose to participate, for their time and creativity, which they generously contributed, in exchange for payment, to make this LoRa dataset a reality:
Alejandra @andromedastrid https://es.fiverr.com/andromedastrid
Angybel M @angybelmonteneg https://es.fiverr.com/angybelmonteneg
Geor Hernandez @geor97 https://es.fiverr.com/geor97
Valeria Perez @valeriap21 https://es.fiverr.com/valeriap21
Whitney M @fabflop https://es.fiverr.com/fabflop
And many other models, without whose participation this would not have been possible. Thank you!
Beyond this, some technical aspects:
In short: it gives a more natural, less powerful effect, and due to the complexity of the hands (guys, we'll get over this someday xD), it's less accurate than the original. I've seen good results between strengths 1.10 and 1.35.
Perhaps the biggest mistake was mine in trying to include the possibility of making the gesture with both hands at the same time (which looks really striking). But oh well, refining the dataset for a single hand is planned for V3.
We have extra smiles, better teeth, and perhaps better eyes with positive bleeding.
This V2 version also offers more variability in different body types (the 'supermodel' effect is controlled) and generally provides a greater degree of realism. It's also able to work better with men, I don't know why, and, yes, let's talk about more things to improve:
There's a certain Caribbean-style bleeding effect and, above all, an excess of tanning in the skin. On the other hand, the feature in close-ups has become very problematic and used to work better. Beyond that, the drawing and anime capabilities have been reduced.
But hey, nothing that can't be fixed for a V3!
Have fun! (And if you want to check out these 21st-century-minded models' Fiverr profiles, even better!)
———
Esta es una revision completa del dataset original, con la novedad de que ahora mismo está íntegramente compuesto con material personalizado creado a medida para este lora trabajando con el arte de modelos reales, de carne y hueso, (con muy poco material sintético, menos del 10%).
En este sentido lo primero es dar crédito a aquellas modelos, que así lo han decidido, por su tiempo y por la creatividad que han cedido, de manera remunerada, para que el dataset de este LoRa sea realidad:
Alejandra @andromedastrid https://es.fiverr.com/andromedastrid
Angybel M @angybelmonteneg https://es.fiverr.com/angybelmonteneg
Geor Hernandez @geor97 https://es.fiverr.com/geor97
Valeria Perez @valeriap21 https://es.fiverr.com/valeriap21
Whitney M @fabflop https://es.fiverr.com/fabflop
Y otras tantas modelos, que sin su participación esto no podría haber sido posible, ¡Gracias!
Más allá de esto, cuestiones técnicas:
En una línea: da un efecto más natural, menos potente y, por la complejidad de las manos (chicos, algún día lo superaremos xD), falla más que el original. Yo he comprobado buenos resultados entre fuerza 1.10 y 1.35.
Quizá el mayor error fue mío al intentar incluir también la posibilidad de hacer el gesto con las dos manos a la vez (que queda super llamativo). Pero bueno, queda pendiente para una V3 depurar el dataset para una sola mano.
Tenemos un extra de sonrisas, mejores dientes y quizá mejores ojos como un sangrado positivo.
Luego también esta versión V2 ofrece más variabilidad de diferentes tipos de cuerpos (se controla el efecto ‘super modelo’) y en general aporta un grado de realismo mayor. También es capaz de trabajar mejor con hombres, no sé por qué, y, eso sí, vamos con las cosas a mejorar:
Se sangra cierto estilo caribeño y, sobre todo, un exceso de moreno en la piel. Y por otro lado sacar el dedo con los primeros planos se ha vuelto muy problemático y antes funcionaba mejor. Más allá de esto, las capacidades con dibujo y anime han quedado reducidas.
Pero bueno, ¡nada que no se pueda intentar arreglar para una V3!
¡A pasarlo bien! (Y si queréis pasaros por el Fiverr de estas modelos con visión para el siglo XXI, mejor que mejor!)
V1
Pretty unstable and maybe fail like half the time. But hey, the usual hands and fingers are a challenge with AI stuff.
Be that as it may, making the middle finger gesture will now go from being something impossible to being something workable with a little time and care thanks to this LoRa.
Have fun!
Bastante inestable y quizá falle como la mitad de las veces. Pero oye, las manos y los dedos de siempre son un reto con cosas de IA.
Sea como fuere, hacer el gesto del dedo ahora pasará de ser algo imposible a ser algo trabajable con un poco de tiempo y mimo gracias a este LoRa.
¡A pasarlo bien!
Description
V2. Dataset redesigned with real models. With a more Caribbean than general effect, but very natural and realistic.
——
V2. Dataset repensado con modelos reales. Con un efecto más caribeño que generalista, pero muy natural y realista.
FAQ
Comments (17)
The cover say: Greetings from President Putin~
¡What? xD
@JoseGomezFreelance It's just a joke. You look at the Mona Lisa on your cover and imagine her with short hair. You'll see the point of my joke. Also, thanks for the buzz bro.
Esto es solo una broma. Mira a la Mona Lisa en tu portada e imagínatela con el pelo corto: verás que se parece mucho a alguien. Entonces comprenderás el significado de mi broma. Además, gracias por el comentario, amigo.
@BIG_A You have a lot of imagination xDD
-
Tienes mucha imaginación xDD
No... ¡ahora veo a Putin en la Mona Lisa! ¡Qué has hecho! xDDDD
-
No... now I see Putin in the Mona Lisa! What have you done! xDDDD
@JoseGomezFreelance LOL~
You didn't even look at your own cover images. Whatever you're doing, whatever your workflow is. Stop.
It's called transparency in a work that's somewhere between experimental research and promotion. I don't know about you, but I hate it when the cover images look amazing and then LoRa only does 20% of the work. My images aren't pre-selected, same seed, same parameters: the success rate is clear, so you know whether you're interested or not. And my recommendation: if you're only interested in power and reliability, choose any of the V1 versions.
And... Nothing and no one can stop me xD
@JoseGomezFreelance No a normal person sees a 50% or higher failure rate and thinks "ah well that was wrong, time to try again" and instead you made a publication page. That's wasting everyone's, including your, time. Why not upload all the checkpoints from the entire run? "for transparency". Oh you didn't. Why? You know why.
@32Bitshifter Clearly we don't play the game the same way xD
And I didn't upload all the checkpoints because I do the work of filtering them, and this one is the best and most balanced :)
The million-dollar question is... what are you doing here with more than 10 times more downloads than me? To share your wisdom altruistically?
@JoseGomezFreelance I think what he means to say, is that if you're trying to promote your model, it's better to put your best foot forward and show what the model is capable in ideal circumstances, rather than include failures in your showcase.
While I respect transparency, your showcase is meant to be a representation of what the model is supposed to do, so end-users know what they should be looking for in a successful result. It's perfectly fine to cherry-pick results for your showcase. Just mention in the description somewhere that results aren't consistent and don't result in a perfect output every time. People will understand, especially with Wan/video gen, since it is still a bit of a "frontier" when it comes to training and what does and doesn't work.
At the end of the day, it's your model to do whatever you want with, but I'm also of the mind that a showcase should be a "showcase". It's not pulling the wool over people's eyes, it's giving their imagination a jumping off point for their own ideas by showing the best of what your model can do.
@anon4324 I understand perfectly. The thing is, ideal conditions often distort the true potential of LoRa (and there, as with photo retouching, there's a gray area, and I've decided not to delve into it).
Anyway, I appreciate your point, anon4324.
The thing is, for my work and personal brand, I'm choosing a different approach :) (And that way, when I nail it, it'll be clear.)
Haven't had much luck actually getting the middle finger, but get wagging their finger a lot. Should the lora be used in both high and low, or only one of them?
It's for wan 5B xD
Okay, I'm obviously not understanding something as I thought Wan 2.2 has high and low models.
@jonk999 And that's true, only that there are 3 different models of Wan 2.2, there is this one, the small one 5B and then the two 14B (hight and low), which are used together :)
@JoseGomezFreelance Thanks for clarifying. Things get so confusing!