aslbeaver / README.md

MidnightRunner

Update README.md

4344259 verified 4 months ago

preview code

raw

history blame contribute delete

7.2 kB

metadata

license: cc-by-nc-4.0
language:
  - en
base_model:
  - MidnightRunner/MDNT_Illus
tags:
  - SDXL
  - text-to-image
  - ComfyUI
  - Automatic1111
  - Diffuser
  - educational
  - sign-language
  - mascot
pipeline_tag: text-to-image
library_name: diffusers
metrics:
  - FID
  - IS
widget:
  - text: >-
      (high-quality cartoon illustration), (vibrant colors), (smooth gradients),
      (polished details), (child-friendly design), aslbeaver, asl_signing
    parameters:
      negative_prompt: (ugly, worst quality, extra limbs, lowres, watermark)
    output:
      url: aslbeaver_e000002_01_20250415205323.png

Model Card for aslbeaver

The aslbeaver LoRA was developed by midnight1111 to generate a cartoon beaver mascot designed to demonstrate ASL (American Sign Language) gestures in educational game environments. The model supports expressions like "phone_on_forehead" for gesture guessing games and animated classroom scenes.

Model Details

Model Description

Model Type: Diffusion-based text-to-image LoRA model built on SDXL

Checkpoint Compatibility: This LoRA works best with SDXL checkpoints such as Illustrious-XL and is compatible with ComfyUI and Automatic1111.

Style & Focus: Designed for expressive cartoon generation focused on educational content, particularly American Sign Language (ASL) gesture storytelling.

Model Name: aslbeaver
Developed by: midnight1111
Finetuned from: https://civitai.com/models/981618?modelVersionId=1405366
Base Model: https://huggingface.co/MidnightRunner/MDNT_Illus
Model Type: LoRA (Low-Rank Adaptation)
License: CC BY-NC 4.0

Model Sources

Demo Use: https://www.playwithasl.com
Repository: https://huggingface.co/MidnightRunner/aslbeaver

Prompt Format

Positive Prompt Example

aslbeaver, phone_on_forehead, gaming_mode, cheerful expression, bright classroom, tail behind

Negative Prompt Example

lowres, worst quality, low quality, bad anatomy, bad hands, 4koma, comic, greyscale, censored, jpeg artifacts, overly saturated, overly vivid, (multiple views:1.1), (bad:1.05), fewer, extra, missing, worst quality, jpeg artifacts, bad quality, watermark, unfinished, displeasing, sepia, sketch, flat color, signature, artistic error, username, scan, (blurry, lowres, worst quality, (low quality:1.1), ugly, (bad anatomy:1.05), artist name, (patreon username:1.2), elongated torso, anime blush,

Settings

Sampling Methods: Euler A, DPM++ 2M Karras, DPM++ 3M SDE
Sampling Steps: 25–30
CFG Scale: 5.0
Clip Skip: 1

Uses

Direct Use

Generate cartoon mascot images for ASL guessing games or classroom visuals.

Downstream Use

Educational apps, gesture-based gameplay designs, comics, posters, signage.

Training Details

  engine: kohya
  resolution: 1024
  epochs: 13
  batch_size: 6
  repeats_per_image: 9
  bucket_enabled: true
  scheduler:
    type: cosine_with_restarts
    cycles: 3

optimizer:
  type: AdamW8Bit

learning_rate:
  unet: 0.0005
  text_encoder: 0.00005

network:
  dim: 32
  alpha: 16

precision:
  use_fp16: true
  noise_offset: 0.1
  min_snr_gamma: 5

clip_skip: 1
keep_tokens: 1
shuffle_caption: true

Inference Sample (Diffusers)

from diffusers import StableDiffusionPipeline
import torch

# Load base model
base_model = "path_to_your_sdxl_or_checkpoint"

pipe = StableDiffusionPipeline.from_pretrained(base_model, torch_dtype=torch.float16)
pipe.to("cuda")

# Load the aslbeaver LoRA
pipe.load_lora("aslbeaver.safetensors", alpha=0.7)

# Define prompt
prompt = "aslbeaver, phone_on_forehead, gaming_mode, cheerful expression, bright classroom, tail behind"
negative_prompt = "(ugly, worst quality, extra limbs, lowres, watermark)"

# Run pipeline
result = pipe(prompt=prompt, negative_prompt=negative_prompt).images[0]
result.save("aslbeaver_result.png")

How to Use in ComfyUI

Load your base SDXL checkpoint in a CheckpointLoaderSimple or similar node.
Use LoraLoader or LoraStack node to load aslbeaver.safetensors.
- Set strength_model to around 0.8.
- Set strength_clip to 1.0 for full text embedding effect.
Use a standard SDXL pipeline with Positive Prompt, Negative Prompt, and Sampler setup as needed.
Render as usual.

How to Use in Automatic1111 (A1111)

Copy aslbeaver.safetensors into:
```
/models/Lora/
```
In the txt2img or img2img interface, enter the LoRA trigger:
```
<lora:aslbeaver:0.8>
```

Combine with your prompt:

aslbeaver, phone_on_forehead, gaming_mode, cheerful expression, bright classroom, tail behind

Add your negative prompt and adjust CFG, steps, etc.

Evaluation

Evaluated visually across all 13 epochs
Best results achieved at epoch 11–13 for styling clarity

Wildcard System: `beaver_wildcards.zip`

This LoRA comes with a complete wildcard tagging system to support flexible, randomized prompt construction. You can use this in both A1111 and ComfyUI environments.

Included Wildcard Files

asl_signing.txt
background_scene.txt
body_proportions.txt
clothing.txt
core_traits.txt
emotion_tags.txt
expression_blocks.txt
expression_pose.txt
gesture_tags.txt
purpose_context.txt
style_rendering.txt

These files contain theme-specific wildcard entries for generating prompts in a modular, randomized fashion. For example, gesture_tags.txt includes ASL-style movements like "wave," "tap head," or "cover mouth."

Example Prompt Syntax

aslbeaver, __gesture_tags__, __clothing__, __emotion_tags__, phone_on_forehead

This syntax will randomly sample one entry from each wildcard file referenced.

Using in ComfyUI

Install: ComfyUI-Wildcards
Path: /ComfyUI/custom_nodes/ComfyUI-Impact-Pack/custom_wildcards/beaver_wildcards/
Place .txt files there and use the Wildcard node to insert __beaver_wildcards/__filename__ pattern.

Using in AUTOMATIC1111 (A1111)

Install: stable-diffusion-webui-wildcards
Path: /extensions/stable-diffusion-webui-wildcards/wildcards/beaver_wildcards/
Then reference wildcards as __beaver_wildcards/filename__ inside your prompt text field.

Wildcarding helps automate image variation for testing, prompt diversity, or simulation of batch.

Environmental Impact

Hardware: 1x RTX 3090 (24GB VRAM)
Runtime: ~12 hours
Emissions: Low

Model Card Authors

midnight1111

Model Card Contact

For support, inquiries, or demo collaboration:

This model is intended to assist in visual communication and cartoon asset creation. Always validate generated ASL signs before public or instructional use.