borozkut commited on
Commit
34d396e
·
verified ·
1 Parent(s): 4e42459

Create README.md

Browse files

📊 Training Info

Epochs: 16

Training Steps: 2,960 total (185 steps per epoch × 16 epochs)

Batch Size: 1 image per step (185 images per epoch after repeats)

Clip Skip: Default (no skipping, uses all CLIP layers)

Min SNR Gamma: 5

Loss Type: L2

LoRA Rank (Dim): 16

Alpha: 1

🎚 Recommended Strength Settings

Min Strength: 0.6 (subtle style influence)

Max Strength: 1.0 (strong comic-style effect)

Default Sweet Spot: 0.75 (balanced detail & style retention)

Files changed (1) hide show
  1. README.md +48 -0
README.md ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ base_model:
3
+ - black-forest-labs/FLUX.1-Krea-dev
4
+ tags:
5
+ - lora
6
+ ---
7
+ Comic-Style LoRA for FLUX.1-dev (c0m4i4)
8
+ Base Model: FLUX.1-dev
9
+ Type: LoRA (Flux Architecture)
10
+ Trigger Word: c0m4i4
11
+
12
+ This LoRA brings a clean, bold comic-style aesthetic to your Flux generations. Trained on 37 high-resolution images at 1024px, it captures strong linework, vibrant color blocking, and stylized shading—perfect for portraits, character art, and vivid scene work.
13
+
14
+ The model was trained for 16 epochs using Flux’s fp8 mixed-precision setup, with LoRA rank 16 for a balance between quality and file size. It adapts well to a variety of prompts, retaining the core comic feel while allowing for subject and scene diversity.
15
+
16
+ Recommended Use:
17
+
18
+ Resolution: 768–1024
19
+
20
+ Sampler: Flow/Euler (Flux schedulers)
21
+
22
+ Steps: 24 (20–28 works well)
23
+
24
+ Guidance (CFG): 1.0
25
+
26
+ LoRA Strength: 0.75 (UNet) / 0.6 (Text Encoder)
27
+
28
+ Negative Prompt: low quality, artifacts, watermark
29
+
30
+ Example Prompts:
31
+
32
+ c0m4i4 as president, bold lines, vibrant shadows, comic illustration
33
+
34
+ c0m4i4 in a black shirt standing in front of a stained glass window, glowing details, clean lines
35
+
36
+ c0m4i4 playing a guitar, warm tones, sharp ink outlines
37
+
38
+ Training Details:
39
+
40
+ Dataset: 37 images × 5 repeats (185 samples per epoch)
41
+
42
+ Network: LoRA rank 16, alpha 1
43
+
44
+ Loss: L2 with min_snr_gamma=5 and noise_offset=0.1
45
+
46
+ Text Encoders: CLIP-L + T5-XXL (mixed bf16/fp8)
47
+
48
+ This LoRA is designed for artists and creators who want a consistent, high-quality comic style without sacrificing detail or prompt flexibility.