Sign In

Retro Comic Flux

285
1.7k
1.3k
150
Verified:
SafeTensor
Type
LoRA
Stats
376
158
Reviews
Published
Nov 8, 2024
Base Model
Flux.1 D
Trigger Words
c0m1c
comic book panel
Hash
AutoV2
AF31BEEE9E
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.

Retro Comic Flux

Retro Comic Flux is a Flux LoRA trained on a dataset consisting entirely of public domain images manually sourced, cropped and enhanced. I used ~50 images 3 repeats 15 epochs and 0.0001 learning rate. These were all captioned using Joy Caption Batch.

This is very experimental and I had a difficult time sourcing high quality images of animals and vehicles in this style, so for now it's primarily good with human figures. It works best when you describe the character and a scene for a backdrop. It struggles with hands and eyes sometimes. If something is really not in the dataset it defaults to an ordinary AI illustration style.

USE DEIS SAMPLER FOR THE BEST RESULTS.

v1: The initial model was trained on lower resolution source images and had limited text handling capabilities.

v2: This version incorporates both the original public domain images and additional high-quality comic book scans made from purchased source public domain material. The training process was optimized to 650 steps based on TensorBoard analytics, finding the sweet spot between convergence and overfitting. The result is a more versatile model with improved text generation and better overall performance.

v2 Usage Tips: When using "halftone" in prompts, include "colorful" to avoid monochrome output. All generated images include ComfyUI workflow data in their metadata - simply click the node icon in the CivitAI gallery to copy the 2-pass generation workflow, then paste directly into ComfyUI to replicate the results.

I used very similar settings as I have with my previous LoRAs so I won't go into detail for the sake of my time and brevity, low learning rates with less repeats and good data work best for simple styles.

I was careful in my selection of images and preprocessed each image in Photoshop to bring out vivid colors, reduce the yellowing and boost contrast. This helped me to maintain a consistent coloring and I believe it improved the training quality. When I did this I also removed the speech from speech bubbles from some of the training data for more flexibility and to help Flux understand context. You can explicitly state "empty speech bubble" in your prompts and it tends to work. If you want to dive in deeper try using an image captioner to describe a style you want to emulate, the images have rich captions and the model responds well to natural language.

If you enjoy this resource please send BUZZ ⚡️so I can keep experimenting and sharing. Please share your images in the gallery after downloading them, by sharing them in the gallery the system gives creators buzz to help with generating and training.

Trigger keyword: 'c0m1c' 'comic book panel'

Recommended strengths: 0.7 - 0.9