geometrics版本v3.0 (ID: 593228)

16-aout-2024

New try with this V4.0

Maybe the last one, because I don't understand anything on what SD understand itself. Will try landscapes now.

Play with the lora's weight, token weight and CFG to have different results. Apparently having asked civitAI to use more steps, geometrics V4 generaly take over the image. Hungrily.

Colors are generaly respected (little bleeding), their relative place also, if a little less, and their geometric form often when not too much are asked. Circle in particular like to take over.

Tokens :

• geometrics (not sure it change anything. it's supposed to be the trigger...)

• COLOR square

• COLOR square with COLOR center

• COLOR background

• COLOR circle

• COLOR circle on top of center

I'm going to put a bounty to see if anybody can explain me what I don't understand...

---

09-jul-2024

After playing with my superheroine cat CinderFluff for no perverted reasons while listening to various eurobeat Discord remixes (you can't judge me), I think I'm on the folowing points :

1. Geometrics is low weighted to the point it can be called feeble. Don't hesitate to weight it to more than 1.

2. gms_ works when it wants, and is lazy. Maybe because gms_ is 4, wich is used as token if what I read is true.

3. I think I perceive some problems in the captioning. It was predictable since it's my first Lora.

4. Geometrics don't answers to modified colors (as "embroidered blue" or "powder grey".

I'm going to work on all that

---

V3. Continuing my tests. Still don't really understand.

Tokens are :

New :

• COLOR circle & COLOR gms_circle

• COLOR background & COLOR gms_background

Old (still working, of course) :

• geometrics

• COLOR square

• COLOR gms_square

• COLOR border

• COLOR gms_border

Still need feedback, please. Could people using it post examples of what they do with it ?

__________________________

V2.5, with all the color combinations (8 training colors on the same 8 colors, square on border), each with its caption file.

Tell me if that change something for you.

But I suddenly wonder if you CAN really reproduce exactly objects like geometric figures, since SD is basically a random sampling machine...

_____________________________

Tokens are :

1. geometrics

2. COLOR square

3. COLOR gms_square

4. COLOR border

5. COLOR gms_border

square /gms_square (same for border) tag the same training images, but "gms_stuff" is not supposed to be known by SD, meaning less blending with what your checkpoint know as "border" or "square" (I'm not sure I understand well, since I don't see a lot of difference, after all).

Images in V2 are also trained in 540*540

Tell me how it's working for you ?

_____________________________

That's my first lora. Pretty simple :

COLOR square, COLOR border

made on CivitAI on this model : aamAnyloraAnimeMixAnime_v1.safetensors

Lora that try to make a square with a border. For reasons of his own, apparently is FAR more imaginative when I try it out on my own web1111.

There, it make images using squares. Beautiful and differents with different checkpoints, but nothing that look like the examples CivitAI show me... :-(

If people can help me, please do.

Also, with some checkpoints, unleash it's love for cats.

描述:

Added :

• background and gms_background (around the figure)

• circle and gms_circle (on top a square if there is)

• Tried to train figures where Square and Border are NOT centered in the image.

训练词语: geometrics,square,gms_square,border,gms_border,circle,gms_circle,background,gms_background