this is for abstract foam-sponge drawing, not for girls, dataset does not contains any "real" objects that has wikipedia article to describe it or words to call it.

Good in img2img, hard to use in txt2img

This SD1.5x version produce more clean output then SDXL version and a bit easy to interaction with (realistic) girls, but anyway very hard to properly activate and settings.

Half of example pictures made with 2-stage workflow in Comfy UI: SDXL + img2img with SD1.5 . This has no reason to use SDXL in this case except experiment and is waste of time and VRAM.

Do not expect good results in first 20 attempts to type prompt. It is difficult to get good output and there were trained over 100 ti, hypernetworks and lora to compare results (comparison failed for now as they are very different and work in different conditions). If you want reproduce result exactly as on picture - write me - i will search json for Comfy Ui and exactly lora, because seems civitai now delete workfkow from metadata

Lexicon / prompt 辞書

activation by these tags working not very good, need to improve training:

v04

ethernalSponge04 - creates sponge like texture good with asteroid, blob, textile, landscape surface

sbelt04 - structure with thick voids between belts. Also capable create belts mess, liana, probably veins but not tested

blobsponge04 - helps to create finite bodies, but not seamless texture.

comboPol04 - create cell-texture and shapes like comb, sea stars. Good with comb, hexagon

snudle04 - noodle (pasta) like elements

ey02 - creepy eye-shaped structure (fill whole picture). I don't know how he settled in this dataset and if this make other things worse

addition tags:

%color - very important, different colors change whole composition

gradient | star sky | simple| yellow | aqua background, - colors of background very important too

sponge, foam,

pores, holes, - for thin or thick holes in surface / 表面の穴

seamless texture,

colorful red border | outline

3d, DOF, blob, asteroid - helps to create finite bodies, but not seamless texture.

net, cage, web, threads, thread - thin version of es11grat grate, usually 1-pixel. web create more round shapes like typical spider web

object, - single or multiple objects that not spread all canvas

multiple objects, - that not spread all canvas

thick, thin, - change texture scale, especially for belts and holes

roots,

portal, orb

prickly, needles, barbed, spiny, cactus,

lattice, fence, grate

polka dot, polka dot background

beige, yellow theme,

architecture, building,

bionic, organic shape

futurism, futuristic,

no humans

Usually avoid tags: corals, sea life, anemone because it is very easy shift SD to draw unstylish semi-natural forms that looks bad.

Negative request

common negative (that you use with girls render) sometimes can prevent this lora to draw anything good. So be very careful, Empty neg or only blur is ok. Need to find what makes it better.

Sometimes better to set negative prompt corals, anemone, food,

cut gems, - add roundness to shapes

Comparison v04 LORA

they are different but espong04Dream8 usually more interesting

Typical workflow in ComfyUI ##TODO##

Settings

Strength: each subject need different, precise and separate control of model and clip strength. 0.2 ... 0.9 is ok.

good with any tested resolution from 512 to 1600

CFG 5...15

Step 25-45, sometimes produce better then very precise steps

About and TODO

if you want to continue training yourself - i can publish dataset as it contains only my own images