Generation Suggestions

As per the training standards I set for SDXL-Simulacrum, Sim-Flux1S, and Sim-Flux1D;

the entire structure is built with a combination of

• plain English captions

• before Booru tags

- for direct integration into many models and reusable utility.

Works if you load it with SDXL in forge as a normal lora.

Load it as a normal lora with any model that has CLIP_L or CLIP_G in Comfy and pass through it's clips.

• LIGHT works best for SDXL-Simulacrum-V2 and SDXL similar models.

• MEDIUM handles derived CLIP_L and CLIP_G models like SD3 pretty well.

• POWERFUL works on Pony, Illustrious, and many heavily altered models.

This is basically my first attempt at a CLIP Lora. It's highly responsive and potent to grids.

Use the same grid_[[letter]row][[number]column] concept as SDXL-SimV2 aka Dumb Flux;

• grid_a1, grid_c5, etc.

Enables more functionality; don't actually put the [ ], replace the text with stuff.

• A 2d grid with five horizontal rows and five vertical columns. There is a total of 25 potential positions for icons.

• [ color ] outline

• a 5x5 grid of [ ... ]

• a 3x3 grid of [ ... ]

• a grid of various shapes

• A 2d icon that fits within a grid

• row of [object shapes] and [colors]

• column of [object shapes] and [colors]

Negative Options -> negative if the grid is TOO STRONG at low strengths

• grid

• outline

• chaotic grid

• noisy shapes, chaotic shapes

A little training background

I have a better one in the works, a much better one with less images. So stay tuned.

Due to not training the UNET, it automatically works on anything with CLIP_L or CLIP_G. If there are any remnant datas or blocks saved into it on the unet-front, you can be certain that they aren't necessary.

Experiment for yourself, you'll see with lower strength entire images are altered due to the selection of images and not using any masking to train. It's intentionally designed to bleed everything in a careful way and provide both structured and non-structured shapes for generation.

The training data is simple geometric grid images and nodes that introduce various grid element controllers trained specifically to the CLIP_L and CLIP_G pair using SDXL-Simulacrum v2 Full in the Civit trainer.

I am trying to apply a larger rule than THE RULE OF 3 to SDXL with this grid concept. The outcome shows promise with all models so far not just SDXL, as now it's using 7 more often than it's using 3 in SDXL for conceptual integration. Main problem now is, due to multires it destroys a lot of positioning information, so this is just an experiment and a tool.

I've included the training data in a zip file. You'll see how simple things like this really are to make.

If you finetune from this, DO NOT use multires noise, as the finished product will lose screen association rapidly.

Training a fully finetuned fairly heavy model like SDXL-Simulacrum with this data will not work en-masse.

Models like this are good for superimposed imprints, but not good for inclusion in core dataset data.

This is a bit of a peek into my tagging style, so if you're interested take a look at the data and the caption files. A peek into the madness of what built SDXL-Simulacrum into what it is now.

A 2d icon that fits within a grid., A singular 2d icon meant to fill an icon placement within a larger grid with strange shapes., A 2d chaotic icon of random shapes., A singular 2d icon with chaotic shapes with a black border.

A 2d icon that fits within a grid., An empty 2d icon placement with an empty background and a colored border and a black outline within the colored border.

A 2d grid with five horizontal rows and five vertical columns. There is a total of 25 potential positions for icons., There are five different colors of shapes filling the five rows.

A 2d icon that fits within a grid., An empty 2d icon placement with an empty grey  background and a grey border.

I had CIVIT tag using it's tagging model. The appended tags are attached to each caption file in the zip folder and are a little different than the original caption data.

It works with many of the models that I've tested; as many of them simply have little to no grid control for characters attached to their CLIPs.

Only real problem is it was trained with MultiRes enabled, which means you can't really finetune off this thing since the entirety of Simv2 was trained on Original noise. Finetuning off multires with sdxl sim v2 full tends to break the grid structure.

2 DIM

128 ALPHA

Due to the low dimensions, it more often superimposes imprints over the image and then loses track of them rapidly on half strength or less. A fantastic outcome from a preliminary reinforcement experiment.

It tends to solidify grids in many places for models.