V1.2 (Anima Civitai):
134-image Lora capable of generating images in the art style of the webcomic Dustbunnies made by Fluf Dustbunny (aka Eric Huelin).
prompts: 2001, 2002, 2003, @flufdustbunny, dustbunnies
optional prompts: flat colors, shaded, shading
negative prompts: text, speech bubble, worst quality, low quality, blurry, artist name, watermark, username, signature, monochrome, greyscale, border, white border, black border, copyright symbol, sketch, 2 comic panel, 3 comic panel, 2koma, 3koma, blood, dirty, multiple girls, multiple boys, multi head, multiple heads, ugly, score_1, score_2, score_3, blurry, jpeg artifacts, chromatic aberration
Engine: AI Toolkit
Base model: Anima Base v1.0
Steps: 268
Epochs: 20
Train Batch Size: 1
Resolution: 1024
Enable Bucket: On
Shuffle Tags: On
Keep Tokens: 2
Clip Skip: 2
Flip Augmentation: Off
Unet LR: 0.0005
Text Encoder LR: 0.00005
LR Scheduler: cosine
LR Scheduler Cycles: 3
Min-SNR Gamma: 5
Network Dim: 8
Network Alpha: 8
Noise Offset: 0.03
Optimizer: AdamW8Bit
V1.1: (Anima Colab):
130-image Lora capable of generating images in the art style of the webcomic Dustbunnies made by Fluf Dustbunny (aka Eric Huelin). Model created in Google Colab.
prompts: 2001, 2002, 2003, @flufdustbunny, dustbunnies
optional prompts: flat colors, shaded, shading
negative prompts: text, speech bubble, worst quality, low quality, blurry, artist name, watermark, username, signature, monochrome, greyscale, border, white border, black border, copyright symbol, sketch, 2 comic panel, 3 comic panel, 2koma, 3koma, blood, dirty, multiple girls, multiple boys, multi head, multiple heads, obscured, score_1, score_2, score_3, blurry, jpeg artifacts, chromatic aberration
Base model: anima-base-v1.0
Resolution: 1024
Repeats: 1
Flip Augmentation: Off
Shuffle Tags: On
Activation Tags: 2
Steps: 804
Epochs: 12
Unet LR: 4e-5
Text Encoder LR: 0.0
LR Scheduler: cosine_with_restarts
LR Scheduler Number: 3
Min-SNR Gamma: 5
Network Dim: 32
Network Alpha: 16
Train Batch Size: 2
Optimizer: AdamW8Bit
V1: (Anima):
128-image Lora capable of generating images in the art style of the webcomic Dustbunnies made by Fluf Dustbunny (aka Eric Huelin).
prompts: 2001, 2002, 2003, @flufdustbunny, dustbunnies
optional prompts: flat colors, shaded, shading
negative prompts: text, speech bubble, worst quality, low quality, blurry, artist name, watermark, username, signature, monochrome, greyscale, border, white border, black border, copyright symbol, sketch, 2 comic panel, 3 comic panel, 2koma, 3koma, blood, dirty, score_1, score_2, score_3,
Base model: Anima Base v1.0
Steps: 2500
Epochs: 20
Train Batch Size: 1
Resolution: 1024
Enable Bucket: On
Shuffle Tags: Off
Keep Tokens: 2
Clip Skip: 2
Flip Augmentation: On
Unet LR: 0.0002
Text Encoder LR: 0
LR Scheduler: constant
LR Scheduler Cycles: 3
Min-SNR Gamma: 0
Network Dim: 32
Network Alpha: 32
Noise Offset: 0
Optimizer: AdamW8Bit
V9.2 (NoobAI Colab):
132-image Lora capable of generating images in the art style of the webcomic Dustbunnies made by Fluf Dustbunny (aka Eric Huelin). Model created in Google Colab.
prompts: 2001, 2002, 2003, by flufdustbunny, dustbunnies
optional prompts: flat colors, shaded, shading
negative prompts: text, speech bubble, worst quality, low quality, blurry, artist name, watermark, username, signature, monochrome, greyscale, border, white border, black border, copyright symbol, sketch, 2 comic panel, 3 comic panel, 2koma, 3koma, blood, dirty, multiple girls, multiple boys, multi head, multiple heads, ugly, jpeg artifacts, chromatic aberration
Base model: NoobAI-XL EPS 1.1
Shuffle Tags: On
Activation Tags: 3
Repeats: 1
Steps: 800
Epochs: 20
Unet LR: 1e-4
Text Encoder LR: 1e-5
LR Scheduler: cosine_with_restarts
LR Scheduler Cycles: 3
LR Warmup Ratio: 0.05
Min-SNR Gamma: 8
IP Noise Gamma: 0.05
Network Dim: 32
Network Alpha: 16
Train Batch Size: 4
Optimizer: AdamW8Bit
V9.1 (NoobAI):
130-image Lora capable of generating images in the art style of the webcomic Dustbunnies made by Fluf Dustbunny (aka Eric Huelin).
prompts: 2001, 2002, 2003, by flufdustbunny, dustbunnies
optional prompts: flat colors, shaded, shading
negative prompts: text, speech bubble, worst quality, low quality, blurry, artist name, watermark, username, signature, monochrome, greyscale, border, white border, black border, copyright symbol, sketch, 2 comic panel, 3 comic panel, 2koma, 3koma, blood, dirty, multiple girls, multiple boys
Base model: NoobAI-XL EPS 1.1
Epochs: 20
Repeats: 7
Train Batch Size: 4
Steps: 4550
Resolution: 1024
Enable Bucket: On
Shuffle Tags: On
Keep Tokens: 2
Clip Skip: 2
Flip Augmentation: On
Unet LR: 0.0005
Text Encoder LR: 0.00005
LR Scheduler: cosine_with_restarts
LR Scheduler Cycles: 3
Min-SNR Gamma: 5
Network Dim: 32
Network Alpha: 16
Noise Offset: 0.1
Optimizer: AdamW8Bit
V9 (NoobAI):
128-image Lora capable of generating images in the art style of the webcomic Dustbunnies made by Fluf Dustbunny (aka Eric Huelin).
prompts: 2001, 2002, 2003, by flufdustbunny, dustbunnies
optional prompts: flat colors, shaded, shading
negative prompts: text, speech bubble, worst quality, low quality, blurry, artist name, watermark, username, signature, monochrome, greyscale, border, white border, black border, copyright symbol, sketch, 2 comic panel, 3 comic panel, 2koma, 3koma, blood, dirty
Base model: NoobAI-XL EPS 1.1
Steps: 4000
Epochs: 20
Train Batch Size: 4
Resolution: 1024
Enable Bucket: On
Shuffle Tags: On
Keep Tokens: 1
Clip Skip: 2
Flip Augmentation: On
Unet LR: 0.0005
Text Encoder LR: 0.00005
LR Scheduler: constant
LR Scheduler Cycles: 1
Min-SNR Gamma: 5
Network Dim: 32
Network Alpha: 16
Noise Offset: 0.1
Optimizer: AdamW8Bit
V8 (NoobAI):
106-image Lora capable of generating images in the art style of the webcomic Dustbunnies made by Fluf Dustbunny (aka Eric Huelin).
prompts: 2001, 2002, 2003, by flufdustbunny, dustbunnies
optional prompts: flat colors, shaded, shading
negative prompts: text, speech bubble, worst quality, low quality, blurry, artist name, watermark, username, signature, monochrome, greyscale, border, white border, black border, copyright symbol, sketch, 2 comic panel, 3 comic panel, 2koma, 3koma, blood
Base model: NoobAI-XL EPS 1.1
Repeats: 14
Steps: 4947
epochs: 20
Unet LR: 0.0001
Text Encoder LR: 0.00001
LR Scheduler: constant
LR Scheduler Cycles: 1
Min-SNR Gamma: 5
Network Dim: 32
Network Alpha: 16
V7 (NoobAI):
100-image Lora capable of generating images in the art style of the webcomic Dustbunnies made by Fluf Dustbunny (aka Eric Huelin).
prompts: 2001, 2002, 2003, by flufdustbunny, dustbunnies
optional prompts: flat colors, shaded, shading
negative prompts: text, speech bubble, worst quality, low quality, blurry, artist name, watermark, username, signature, monochrome, greyscale, border, white border, black border, copyright symbol, sketch, 2 comic panel, 3 comic panel, 2koma, 3koma, blood
steps: 4000
repeats: 12
epochs: 20
Unet LR: 0.0001
Text Encoder LR: 0.00001
LR Scheduler: constant
LR Scheduler Cycles: 1
Network Dim: 64
Network Alpha: 32
V6 (NoobAI):
85-image Lora capable of generating images in the art style of the webcomic Dustbunnies made by Fluf Dustbunny (aka Eric Huelin).
prompts: 2001, 2002, 2003, by flufdustbunny, dustbunnies
optional prompts: flat colors, shaded, shading
negative prompts: text, speech bubble, worst quality, low quality, blurry, artist name, watermark, username, signature, monochrome, greyscale, border, white border, black border, copyright symbol, sketch, 2 comic panel, 3 comic panel, 2koma, 3koma, blood
steps: 3400
epochs: 20
Unet LR: 0.0001
Text Encoder LR: 0.00001
LR Scheduler: constant
LR Scheduler Cycles: 1
Network Dim: 64
Network Alpha: 32
V5 (NoobAI):
65-image Lora capable of generating images in the art style of the webcomic Dustbunnies made by Fluf Dustbunny (aka Eric Huelin).
prompts: 2001, 2002, 2003, by flufdustbunny, dustbunnies
optional prompts: flat colors, shaded, shading
negative prompts: text, speech bubble, worst quality, low quality, blurry, artist name, watermark, username, signature, monochrome, greyscale, border, white border, black border, copyright symbol, sketch, 2 comic panel, 3 comic panel, 2koma, 3koma
V4 (NoobAI):
46-image Lora capable of generating images in the art style of the webcomic Dustbunnies made by Fluf Dustbunny (aka Eric Huelin).
prompts: 2001, 2002, 2003, by flufdustbunny, dustbunnies
negative prompts: text, speech bubble, worst quality, low quality, blurry, artist name, watermark, username, signature, monochrome, greyscale, border, white border, black border, copyright symbol, sketch, 2 comic panel, 3 comic panel, 2koma, 3koma
V3 (NoobAI):
34-image Lora capable of generating images in the art style of the webcomic Dustbunnies made by Fluf Dustbunny (aka Eric Huelin).
prompts: 2001, 2002, 2003, by flufdustbunny, dustbunnies
negative prompts: text, speech bubble, worst quality, low quality, blurry, artist name, watermark, username, signature, monochrome, greyscale, border, white border, black border, copyright symbol, sketch, 2 comic panel, 2koma
V2 (NoobAI):
24-image Lora capable of generating images in the art style of the webcomic Dustbunnies made by Fluf Dustbunny (aka Eric Huelin).
prompts: 2001, 2002, 2003, by flufdustbunny, dustbunnies
negative prompts: text, speech bubble, worst quality, low quality, blurry, artist name, watername, username, signature, monochrome, greyscale, border, white border, black border, copyright symbol, sketch,2 comic panel, 2koma
V1 (NoobAI):
11-image Lora capable of generating images in the art style of the webcomic Dustbunnies made by Fluf Dustbunny (aka Eric Huelin).
prompts: 2001, 2002, 2003, by flufdustbunny, dustbunnies
negative prompts: text, speech bubble, worst quality, low quality, blurry, artist name, watername, username, signature, monochrome, greyscale, border, white border, black border, copyright symbol, sketch,
Description
A mini update in which I returned to Civitai to continue training Anima-based Lora models using the AI Toolkit. Here, I chose to reduce the Network Dim and Network Alpha to 8, changed the LR Scheduler to cosine, increased the UNet LR and Text Encoder LR to 0.0005 and 0.00005 respectively, set Flip Augmentation to Off, and reduced the Steps to 268. The resulting model version is a bit closer to Eric Huelin’s art style, but it’s still far from delivering the results I was hoping for, unlike the NoobAI version trained with Kohya. So I’ll continue to refine my training of Anima-based models using the AI Toolkit.