Sign In

(un)serious Musk lora

8
71
0
Type
LoRA
Stats
71
Reviews
Published
Dec 20, 2024
Base Model
Flux.1 D
Trigger Words
A candid photograph of m6zkk
m6zkk
Hash
AutoV2
BFADF406D6
default creator card background decoration
mrnatural's Avatar
mrnatural
The FLUX.1 [dev] Model is licensed by Black Forest Labs. Inc. under the FLUX.1 [dev] Non-Commercial License. Copyright Black Forest Labs. Inc.
IN NO EVENT SHALL BLACK FOREST LABS, INC. BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.

Recommendations and models used in sample images:
Model used is the base blackforest flux1-dev.sft .

Lora Strength 1.20

Ksampler use 30 or 40 steps, with Euler sampler, beta scheduler

flux guidance 3.5 and sometimes up to 5.0
Adding realism flux lora helps to get a better picture. I use it often.

Inpaint

Doesn't work great with flux fill, but if you really need it, use a lora_strength of 3.0 and flux_guidance to 50.0 instead of 30.0 (The third version will be better for inpainting, because I finally find how to resolve this :))

Prompting
Use m6zkk as trigger. if you get an Asian guy, use Caucasian m6zkk (What would Kanye West think about this? :/)
"A candid photograph of m6zkk ..." usually gives good results (it's undoubtedly related to how I prompted the dataset. My bad).

The first version was a proof of concept; I was curious to see what could be achieved from just one high-quality image (actually two: the original and the same mirrored). The result wasn't too bad, but (as tested by some of you) the likeness was relatively inconsistent. Expecting 20% of acceptable/satisfactory generations at best.

For the second version, I used 24 very high-quality images at 1024x1024, all taken between 2022 and 2024 to maintain good consistency across the dataset and preserve the essence of the subject in its contemporary context (the 'old' musk). The lora was trained with the following parameters: 3 repeat 16 epochs on 24 images, totaling 1152 training steps. I also changed the trigger word (from 'm000sk' to 'm6zkk' because in an intermediate version, foam moos was sometimes growing onto his face haha).

Side note: I hate this man with all my soul, but I needed an easily identifiable and recognizable topic for everyone to move forward with my experiments. Given the current events, I think it's an ideal subject for meme... so here's the lora, and it's up to you to do the rest of the work (sharing is caring, keep your buzz and instead share your images in the gallery or through a link in the comments if your creations are incompatible with Civitai's TOS). Have a happy new year and don't forget to have fun! ;)