Verified: 3 months ago
SafeTensor
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.
Creature Shock Flux
My Creature Shock Flux LoRA was trained on approximately 60 images to excel at generating uniquely strange creatures with distinctive features such as fur, sharp teeth, skin details and detailed eyes. While Flux already produces creature images, this LoRA greatly enhances detail, creating more realistic textures like scaly skin and an overall production-quality appearance, making the creatures look truly alive. This one is a lot of a fun and it can do more than you think, prompt adherence is pretty decent, I've included some more details below.
I utilized the Lion optimizer option in Kohya, which proved effective in refining the concept and style without overtraining. The training process involved a batch size of 2, 60 images (no repeats), a maximum of 3000 steps, 35 epochs and a learning rate of 0.0003. The entire training took approximately 4 hours. Images were captioned using Joy Caption Batch, and the model was trained with Kohya and tested in ComfyUI.
The gallery will feature examples with workflows attached, I'm running a very simple 2-pass workflow for most of these, drag and drop the first image into ComfyUI to see the workflow. (It's being analyzed right now, may take a few hours to show up past the filter.)
There are a couple of things with variety that I'd like to improve. I'm still putting the model through its paces, and you can expect v1, trained with some of the generated outputs from v0, to drop soon. I really wanted to share this because I think we, as a community, often get stuck just repeating the same 'recommended' settings without experimenting with how different approaches can break away from default behaviors.
Version Notes:
v0 - This version of the LoRA skews towards a very specific look, it works well but skewed towards heavily textured skin.
v1 - This version was trained with a different process that captured higher quality details, lots more flexibility in the outputs. This is really good at generalizing and can be used in far more creative ways in my testing. Similar dataset but a very different approach for training.
Trigger Words:
cre4tur3
Recommended Strength: 0.7–0.9
Recommended Samplers: heun, dpmpp_2m
renderartist.com