Andres — home

CATEGORY · MONTH YEAR · X MIN

90% Rule

The line between utilising AI to ehnance your brand and your brand becoming syonmuyous with AI slop is somewhere in betwen this undefined area. 

When do you know a gneriton from a model/models is past the piint whe you can just spam genraitons to try and get the 100% vision? 

i call it the 90% rule. that is when i,based on expernance i gained in image prodcution piplines and depciplies like retoucing and time managemtn, and thenmeasureed agiant my experiance with thes models and their archtecuture. to make the deicson to lock a genration. this is what is called human taste, jdgement. its a skilset boudned by huan expiernace, it is something an ai cannot replicate convinginly. its goal is to satirify its goals shaped by training and trainding data. that is often very convinging, but empty image which often satisify algorhitms, not humans that expect authneticity from the brands love and trust.

This is often the case with complex scenes, where scene geometry sits behind more geometry and there's a layering effect, e.g. an interior with many walls and furniture vs a still life with a single plinth and an object on top.

Here is some of my thinking about this concept

I ofnte get to this 90% , having circled and iterated best i can , and then run the prmot once more acros the most sucuesfuk models i used thooguthe interation proess. somtimes i run it 4-5 itmes loking for the unicorn 100% genraitn. rarely hapepns, based on all I' e elanrnt, there is always clean up of some sort, this is the line i've found between slop and curated buy a human, for human eyes. 

Below are some exaples of 90% prmots frommy ok. proejct. once the prmot is 90%, the model issamplying from very similir aprmetns, so the reuslt are often very similir.  

3D · Blender
AI Image Model · ChatGPT Image 2 - Medium 

With this iamge, i was looking for dymanic lgihting and poses, that the subjects were varied and the overall sceen was far enoguh away from slop symmetry. runing the same genritonw ith the sme inpout files and prmot sually genrtes similir results due to the models smapling the same latent space. so balancing the core reuimnets as a checkist and rruning a prmot vs adjusting th e prmot is someting that ive found as a very valuable skill. 

Waht am i loknfor? things like what is easy or ahrd to inpaint, waht can i do msyself with tradional retouhcign.
based on my epierance woking with models, ie found ceain mdoels are beter or worse at inpainting some things over others, and often times its th ehero objets or suject that is the most critical to objverse in these 90% prmots. 

3D · Blender
AI Image Model · MODEL NAME
Above — the 90% result and the final edited image.

Here is some of my thinking about this concept


There are other factors as well, inpaitnig however, small or isoalted, ie you wnat to fix a hand in a iamge in a group shot, introduces destrcutive arcitfacts to the image each time it fed in, hat jsut the nature of a duffuciosn model, and the tech undlyig it,controlling, is in intself, probabilistic, which means each time, it is beign rerrawn, i alway wsugest zooming intot hte scene toamek sure the resutls are not too drastic. 
PLACEHOLDER — reference set, boxed.


thre are other ways t mitigate lighting, colour and composition drift when inpaitng tied direcly to the prmot itself such as strucurra the negative fields at the bottom of the prmot which allows the model is reason {artivle with reasoning models} thrgouh the most imrptjat cerative,positive sections of the prmot without being interuped with negtives at the stat or mixed with posstive that may cause confusion.


then there is cost


which i speka aobut further in this blog add link to blog aobut dmoel rpcing this si somethin to wegiht if, and when you need to genretn umltple isngle apsses on a fin la iamhge, and get lsot in all the genraiotns being made, odnwlaod dropped into the heor iamge and ptainted in or overayed. 
PLACEHOLDER — structure, list or short quoted block.

1. one
2. two
3. three

setup for success


 how best to seup for scuess, work with a chatbot,preferrably the smratest aiable on your paid plan, and ultiamise features like caldues preojcts which are self contianed spaces that have all the functnalirty of caldue, but dont have memoeries and context from your other hats about cats or baking or how to do a deadlift being incldued into its memory for each echat. you can setup the brief and rules that help guie hr mdoel. 

second - experment with difenrnt models with a rpmot that is diverse and clangngeing enough that you can have genraiotns from 5 difenrnt mdels and deide which tsue. a node wokrflow wokrd great here, you can uplaod a sinele inout ref and or text ptmot , and conect it omultple models and run all at once. feels like a prchestrator - beae i guess in some ways it is.

if you have a chacter story, its always god to genroing test runs of chacter sheets that fit uour desitions {link aricibel to lileness chesks when certing mirage charaters}

a chatbot is great at heling to refine prmits, not looka t iamges. i tlka mroe aobut what ive eoarnt about comouter vision (an ai eyes) and its limitations. one fo those is to debir the looka nd feel of a iamge or video. something entirely human. {arcitle on model and human physocalogy}. 

this this ppint, i think yuor iterative progress wshould be enoguh progress with 90% iamgery or hand off to a opoerater that can help get you to the 90%. 


Next Article>>