Suggested by @BigGunsNeverTire
It makes thing in yellow-black, robust, blocky, modular design.
V2.0:
Created exactly as described in Tutorial: konyconi-style LoRA, with no changes.
V1.0:
Created as described in Tutorial: konyconi-style LoRA, with following changes:
1] I've added a few real (non-generated) pictures of excavators -- generators do not get them right.
2] set dimension (rank) to 16 and alpha to 1 --- just for science.
It was trained on 86 pics in total.
Perpetual coffee request: I signed up for Buy-Me-a-Coffee and Ko-fi, mainly because I thought it would be hilarious if someone as artistically untalented as myself received a coin for something creative. It's already happened, and it made my day. AI made the impossible possible :D. If you'd like to add to the amusement, please consider buying me more coffee through the following links:
Description
FAQ
Comments (14)
I will buy that coffee machine ! :)
Very cool idea. Thank you
everything looks better in construction orange
Awesome. It's being great seeing you starting to experiment. I have been finetunning for 9 months or so and I can tell it's the best way to keep you active, not bored and always learning new things. IMO the way to go.
Being said that, the Dim/Rank factoring is extremely complex to explain (and technical) and correlate to epoch/steps/repetitions (keycore of LORA) but think of it as a multiplier factor. I will make some free public post in my patreon with some insight and images but think of this as "if you train a suitcase" (example), the 128 will exaggerate the suitcasing effect and the 8, 16, 32, etc. will lower that exaggeration. There is also the point of with lower DIM/Rank you can go higher CFG and so on and control influence with rank too... so.. it's all dependant to the keycore config and IMO even more dependant to the input image dataset.
Keep experimenting!
I would love to see a version of this with dim/alpha back at the usual settings... I love this aesthetic but it definitely comes through a bit less bold and distinctive compared to previous ones.
You are right, I can see it as well. Currently, I am working on GlowingRunesAI v2.0. I have added the color red, but with these altered parameters, it appears somewhat ordinary compared to v1.0. Once I determine the cause, I will revisit this LoRA.
@konyconi Thanks. In my experimenting with my own "concept" LoRAs, concept adherence drops noticeably at lower dim. I think low dim can work for characters/people but maybe less so for "ideas".
Same here. As an experiment he could do comparisons between 16/16 vs 128/128 for example. Alpha variation when lowering dimensions also affects. @konyconi have you think on releasing a dataset with txt already tagged so people can experiment with? (Will be a good addon to the tutorial).
@LDWorksDavid OK, I will do that.
@konyconi much appreciated!
Now I can claim all those hours playing Satisfactory and Hardspace: Shipbreaker were for inspiration.
Make generation into a business
Write Satisfactory off as a business expense
???
Profit
I love all of your imaginative work!! How about a slime model? Nickelodeon slime lab style :) I'm doing a kid's book and could use serious factory slime :)
Ok. I'll try.
@konyconi AHHH! I just got on and saw the slimeAI - hadn't even seen your response here yet. THANK YOU so much! :) That's awesome! :)
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.