This Pepe FLUX Lora was trained on the same dataset of 569 image caption pairs that I used to train the BS-Pepe-V1 SDXL LoRA. It was trained with a batch size 4 and grad accum 4 using the AdamW optimizer for 4000 steps with a network rank of 64. Training took about 14 hours. Most of the tags from the SDXL version work, some are a little weird. 3D Render, Digital Art, Painting, Girl, CCTV styles all trained very well. Anime, Cartoon and Comic can be a little tricky. Try using the style tags at the end of your prompt.
Description
Tags need to be explored to find the best ones, initial testing indicating that all though everything works, some work better than others. Try using style tags at the end of your prompt.
Training settings:
train:
batch_size: 4
steps: 4000
gradient_accumulation_steps: 4
train_unet: true
train_text_encoder: false
content_or_style: balanced
gradient_checkpointing: true
noise_scheduler: flowmatch
optimizer: adamw8bit
lr: 0.0004
ema_config:
use_ema: true
ema_decay: 0.99
dtype: bf16
FAQ
Comments (8)
Finally, thx!
With what models works?
flux dev
not sure if its the flux rendering style but some of those samples almost look like they were trained on those bing dall-e memes you find on 4chan? not that its a bad thing
That is where the data set came from, it is 500+ dalle3 pepes, i am workin on an improved version of this model.
@blairesilver13 It's really good, it is exactly the styling I wanted. Good work.
Can this be used in Comfy UI?
yes, all demo images were made with ComfyUI
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.