Verified: a year 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.
v2
This LORA seems to reproduce the artist style better when the prompt includes "girl" instead of "anime."
It's difficult to set my own trigger word because the current FLUX fine tuning in sd-scripts doesn't allow for TE training.
As a workaround, I had been using "anime" as the trigger word, but it wasn't always accurate.
This LORA since the dataset contains many images of girls, using "girl" as the trigger word seems to yield better results.
Words like "woman" or "female" are unlikely to work as trigger words, so in those cases, I think "anime" would be a better option.
I have done additional training on v1 using a 832×1216 resolution dataset.
The dataset is identical to v1, except for the resolution.
832x1216, 30images.
No caption.
sd-scripts : fine tuning
optimizer : adafactor ("scale_parameter=False" "relative_step=False" "warmup_init=False")
scheduler : constant
LR : 1e-6
batch size : 6
epochs : 150 (First 100 epochs and additional 50 epochs due to lack of training.)
repeats : 1
v1
This LORA will not generate a good quality image with short prompts.
I would recommend using an LLM to generate longer prompts.
416 x 608, 30 images.
No caption.
sd-scripts : fine tuning
optimizer : adafactor ("scale_parameter=False" "relative_step=False" "warmup_init=False")
scheduler : constant with warmup (100%)
LR : 6e-6
batch size : 30
epochs : 300
repeats : 1