|
--- |
|
license: apache-2.0 |
|
datasets: |
|
- neph1/stable-diffusion-prompt-pairs |
|
- Gustavosta/Stable-Diffusion-Prompts |
|
finetuned-from: |
|
- acrastt/Griffin-3B |
|
--- |
|
|
|
# Model Card for Model ID |
|
|
|
This is an experimental model that translates natural language into prompt tags for stable diffusion. |
|
|
|
The input is synthetic, and generated with chronos-hermes-13b-v2.Q4_K_M |
|
|
|
The tags come from Gustavosta's MagicPrompt-Stable-Diffusion. |
|
|
|
[](https://ko-fi.com/T6T3S8VXY) |
|
|
|
Example (validation set): |
|
|
|
Input: "Render a realistic 3D sneaker with a futuristic design, featuring intricate detail and a hexagon mesh pattern. Use soft, natural lighting to highlight the purple and blue leather, as well as the shiny plastic accents. The sneaker should be viewed from a side angle in close-up, with a sharp focus and neon glowing lines. The final image should be trending on popular art platforms like ArtStation and DeviantArt, and should resemble popular brands like Nike, Asics, and Salomon." |
|
|
|
Output: "realistic 3d sneaker futuristic design by greg rutkowski with intricate detail in the style of alphonse mucha using soft natural lighting to highlight purple blue leather shiny plastic accents closeup sharp focus neon glowing lines trending on artstation deviantart nike asics salomon" |
|
|
|
Ground truth: "realistic 3 d render of a futuristic sneaker, beautiful studio lighting, soft, sharp focus, neon glowing lines, intricate detail, purple and blue leather, soft white rubber, shiny plastic, hexagon mesh, gold filigree, octane render, side view, close up, trending on artstation, deviantart, nike, asics, salomon" |
|
|
|
|
|
Other example (maybe polluted): |
|
|
|
Input: "Cute, friendly barista in Tokyo by Charlie Bowater and Titian and ArtGerm. Intricate face with symmetrical eyes. Japanese cafe setting. Elegant and beautiful with highly detailed elements. Dramatic lighting and sharp focus. Trending on ArtStation with tags such as 'ArtStationHD', 'ArtStationHQ', 'Unreal Engine', '4K', and '8K'. Instruct the artist to create a piece that captures the essence of the barista's personality and the unique atmosphere of the cafe, while showcasing their skill in creating intricate and highly detailed artwork. Use vibrant colors and dynamic lighting to create a visually stunning piece that will stand out on ArtStation and other online platforms." |
|
|
|
Output: "Cute friendly barista by Charlie Bowater Titian artgerm juliette lewis symmetrical eyes Japanese cafe setting elegant beautiful highly detailed intricate, dramatic lighting sharp focus trending on artstation hq unreal engine 4 k 8 k, ArtStationHD ,ArtStationHQ" |
|
|
|
### Model Description |
|
|
|
Example comparisons: |
|
In all cases: 1. Natural language as prompt. 2. Tags. 3. sd-seer output as prompt. 4. sd-seer output, but sdxl |
|
 |
|
|
|
Prompt: Render a realistic 3D sneaker with a futuristic design, featuring intricate detail and a hexagon mesh pattern. Use soft, natural lighting to highlight the purple and blue leather, as well as the shiny plastic accents. The sneaker should be viewed from a side angle in close-up, with a sharp focus and neon glowing lines. The final image should be trending on popular art platforms like ArtStation and DeviantArt, and should resemble popular brands like Nike, Asics, and Salomon. |
|
|
|
 |
|
|
|
Prompt: A whirlwind of souls rushing inside the metaverse, with half-body figures wearing a tiara and a tattoo of a pharaoh on their skin. The android or cyborg character should have a cyberpunk face, and the artwork should be inspired by the styles of Loish, D&D, fantasy, intricate, elegant, highly detailed, colorful, and vivid color. The final piece should be a digital painting that showcases the artist's skill and creativity, and can be submitted to ArtStation or featured in a concept art gallery. |
|
|
|
 |
|
|
|
Prompt: Cute, friendly barista in Tokyo by Charlie Bowater and Titian and ArtGerm. Intricate face with symmetrical eyes. Japanese cafe setting. Elegant and beautiful with highly detailed elements. Dramatic lighting and sharp focus. Trending on ArtStation with tags such as 'ArtStationHD', 'ArtStationHQ', 'Unreal Engine', '4K', and '8K'. Instruct the artist to create a piece that captures the essence of the barista's personality and the unique atmosphere of the cafe, while showcasing their skill in creating intricate and highly detailed artwork. Use vibrant colors and dynamic lighting to create a visually stunning piece that will stand out on ArtStation and other online platforms. |
|
|
|
- **Developed by:** Me |
|
- **Model type:** Llama2 |
|
- **License:** Apache 2.0 |
|
- **Finetuned from model:** https://huggingface.co/acrastt/Griffin-3B |
|
|
|
|
|
## How to Get Started with the Model |
|
|
|
System prompt: "This is a conversation between User and the helpful AI-bot Llama. Llama breaks down image descriptions into precise comma-separated single-word tags adding tags where they might increase quality. Llama writes only in comma-separated tags without asking questions."" |
|
|
|
Prompt template: |
|
|
|
### HUMAN: |
|
prompt |
|
|
|
### RESPONSE: |
|
<leave a newline for the model to answer> |
|
|
|
Temperature: 0.3 |
|
|
|
## Training Details |
|
|
|
### Training Procedure |
|
|
|
I based my dataset on the prompts from this dataset: |
|
https://huggingface.co/Gustavosta/MagicPrompt-Stable-Diffusion |
|
|
|
Based on that, I asked chronos-hermes-v2-13b to generate natural language descriptions. |
|
|
|
I then trained a model on the pairs, with the NL as input and the prompt tags as output |
|
|
|
#### Training Hyperparameters |
|
|
|
# Number of training epochs |
|
num_train_epochs = 5 |
|
|
|
# Enable fp16/bf16 training (set bf16 to True with an A100) |
|
fp16 = False |
|
bf16 = True |
|
|
|
# Batch size per GPU for training |
|
per_device_train_batch_size = 8 |
|
|
|
# Batch size per GPU for evaluation |
|
per_device_eval_batch_size = 8 |
|
|
|
# Number of update steps to accumulate the gradients for |
|
gradient_accumulation_steps = 1 |
|
|
|
# Enable gradient checkpointing |
|
gradient_checkpointing = True |
|
|
|
# Maximum gradient normal (gradient clipping) |
|
max_grad_norm = 0.3 |
|
|
|
# Initial learning rate (AdamW optimizer) |
|
learning_rate = 3e-5 |
|
|
|
# Weight decay to apply to all layers except bias/LayerNorm weights |
|
weight_decay = 0.001 |
|
|
|
# Optimizer to use |
|
optim = "paged_adamw_32bit" |
|
|
|
# Learning rate schedule |
|
lr_scheduler_type = "cosine" |
|
|
|
# Number of training steps (overrides num_train_epochs) |
|
max_steps = -1 |
|
|
|
# Ratio of steps for a linear warmup (from 0 to learning rate) |
|
warmup_ratio = 0.03 |
|
|
|
## Environmental Impact |
|
|
|
- **Hardware Type:** RTX 3060 |
|
- **Hours used:** 6 hours |
|
- **Carbon Emitted:** None? As I use a non-co2 emitting energy provider. |
