PyPotteryInk Model Zoo
Overview
This page lists the available pre-trained models for PyPotteryInk. Each model is show specific pros and cons. You can use these models directly or fine-tune them for your specific needs.
Benchmark image
The benchmark image contains a variety of images and styles for testing each model. It can be used to quickly assess the quality of the output. As the library will be used for a variety of different styles and morphologies, new images will be added to the benchmark images.
Available Models
10k
Description: The base model trained on a diverse dataset of 492 archaeological drawings from multiple contexts including Casinalbo, Montale, Monte Croce Guardia, and Monte Cimino sites.
- Training Data: 492 paired drawings
- Resolution: 512×512 pixels
- Training Steps: 10,000
- Best For: General purpose archaeological drawing inking, starting point for fine-tuning
Parameters used for training:
| Parameter | 10k |
|---|---|
| adam_beta1 | 0.9 |
| adam_beta2 | 0.999 |
| adam_epsilon | 0.00000001 |
| adam_weight_decay | 0.01 |
| allow_tf32 | false |
| checkpointing_steps | 500 |
| dataloader_num_workers | 0 |
| enable_xformers_memory_efficient_attention | true |
| eval_freq | 100 |
| gan_disc_type | “vagan_clip” |
| gan_loss_type | “multilevel_sigmoid_s” |
| gradient_accumulation_steps | 1 |
| gradient_checkpointing | false |
| lambda_clipsim | 5 |
| lambda_gan | 0.5 |
| lambda_l2 | 1 |
| lambda_lpips | 5 |
| learning_rate | 0.000005 |
| lora_rank_unet | 8 |
| lora_rank_vae | 4 |
| lr_num_cycles | 1 |
| lr_power | 1 |
| lr_scheduler | “constant” |
| lr_warmup_steps | 500 |
| max_grad_norm | 1 |
| max_train_steps | 10,000 |
| mixed_precision | “no” |
| num_samples_eval | 100 |
| num_training_epochs | 1 |
| pretrained_model_name_or_path | “stabilityai/sd-turbo” |
| resolution | 512 |
| set_grads_to_none | false |
| test_image_prep | “resized_crop_512” |
| track_val_fid | true |
| train_batch_size | 2 |
| train_image_prep | “resized_crop_512” |
| viz_freq | 25 |
Author: Lorenzo Cardarelli
Benchmark
6h-MCG
Description: Fine-tuned model specialized for Monte Croce Guardia style pottery drawings. Optimized for high-detail preservation and consistent stippling patterns.
- Training Data: 9 paired drawings from Monte Croce Guardia
- Base Model: 10k Base Model
- Resolution: 512×512 pixels
- Training Steps: 600
- Best For: Middle Bronze Age / Recent Bronze Age pottery
Parameters used for training:
| Parameter | 6h-MCG |
|---|---|
| adam_beta1 | 0.9 |
| adam_beta2 | 0.999 |
| adam_epsilon | 0.00000001 |
| adam_weight_decay | 0.01 |
| allow_tf32 | false |
| checkpointing_steps | 100 |
| dataloader_num_workers | 4 |
| enable_xformers_memory_efficient_attention | true |
| eval_freq | 100 |
| gan_disc_type | “vagan_clip” |
| gan_loss_type | “multilevel_sigmoid_s” |
| gradient_accumulation_steps | 4 |
| gradient_checkpointing | false |
| lambda_clipsim | 3 |
| lambda_gan | 0.9 |
| lambda_l2 | 1 |
| lambda_lpips | 10 |
| learning_rate | 0.00001 |
| lora_rank_unet | 128 |
| lora_rank_vae | 48 |
| lr_num_cycles | 1 |
| lr_power | 1 |
| lr_scheduler | “constant” |
| lr_warmup_steps | 200 |
| max_grad_norm | 1 |
| max_train_steps | 600 |
| mixed_precision | “no” |
| num_samples_eval | 100 |
| num_training_epochs | 1 |
| pretrained_model_name_or_path | ./10k.pkl |
| resolution | 512 |
| set_grads_to_none | false |
| test_image_prep | “resized_crop_512” |
| track_val_fid | true |
| train_batch_size | 1 |
| train_image_prep | “resized_crop_512” |
| viz_freq | 50 |
Author: Lorenzo Cardarelli
Benchmark
6h-MC
Description: Fine-tuned model specialized for Monte Cimino style pottery drawings. Optimized for high-detail preservation and engraved decoration.
- Training Data: 15 paired drawings from Monte Cimino
- Base Model: 10k Base Model
- Resolution: 512×512 pixels
- Training Steps: 600
- Best For: Recent Bronze Age pottery / Final Bronze Age / Historic Age
Parameters used for training:
| Parameter | 6h-MC |
|---|---|
| adam_beta1 | 0.9 |
| adam_beta2 | 0.999 |
| adam_epsilon | 0.00000001 |
| adam_weight_decay | 0.01 |
| allow_tf32 | false |
| checkpointing_steps | 100 |
| dataloader_num_workers | 4 |
| enable_xformers_memory_efficient_attention | true |
| eval_freq | 100 |
| gan_disc_type | “vagan_clip” |
| gan_loss_type | “multilevel_sigmoid_s” |
| gradient_accumulation_steps | 4 |
| gradient_checkpointing | false |
| lambda_clipsim | 3 |
| lambda_gan | 0.9 |
| lambda_l2 | 1 |
| lambda_lpips | 10 |
| learning_rate | 0.00001 |
| lora_rank_unet | 256 |
| lora_rank_vae | 48 |
| lr_num_cycles | 1 |
| lr_power | 1 |
| lr_scheduler | “cosine” |
| lr_warmup_steps | 200 |
| max_grad_norm | 1 |
| max_train_steps | 600 |
| mixed_precision | “no” |
| num_samples_eval | 100 |
| num_training_epochs | 1 |
| pretrained_model_name_or_path | ./10k.pkl |
| resolution | 512 |
| set_grads_to_none | false |
| test_image_prep | “resized_crop_512” |
| track_val_fid | true |
| train_batch_size | 1 |
| train_image_prep | “resized_crop_512” |
| viz_freq | 50 |
Author: Lorenzo Cardarelli
Benchmark
4h-PAINT
Description: Fine-tuned model specialized for painted decoration and historical pottery.
- Training Data: 15 paired drawings from Late Bronze Age / Early Iron Age of Southern Italy
- Base Model: 6h-MC
- Resolution: 512×512 pixels
- Training Steps: 400
- Best For: Historic Age / Painted Pottery (No shadows!)
Parameters used for training:
| Parameter | 4h-PAINT |
|---|---|
| adam_beta1 | 0.9 |
| adam_beta2 | 0.999 |
| adam_epsilon | 0.00000001 |
| adam_weight_decay | 0.01 |
| allow_tf32 | false |
| checkpointing_steps | 100 |
| dataloader_num_workers | 4 |
| enable_xformers_memory_efficient_attention | true |
| eval_freq | 100 |
| gan_disc_type | “vagan_clip” |
| gan_loss_type | “multilevel_sigmoid_s” |
| gradient_accumulation_steps | 4 |
| gradient_checkpointing | false |
| lambda_clipsim | 3 |
| lambda_gan | 0.9 |
| lambda_l2 | 1 |
| lambda_lpips | 10 |
| learning_rate | 0.00001 |
| lora_rank_unet | 256 |
| lora_rank_vae | 48 |
| lr_num_cycles | 1 |
| lr_power | 1 |
| lr_scheduler | “cosine” |
| lr_warmup_steps | 200 |
| max_grad_norm | 1 |
| max_train_steps | 400 |
| mixed_precision | “no” |
| num_samples_eval | 100 |
| num_training_epochs | 1 |
| pretrained_model_name_or_path | ./6h-MC.pkl |
| resolution | 512 |
| set_grads_to_none | false |
| test_image_prep | “resized_crop_512” |
| track_val_fid | true |
| train_batch_size | 1 |
| train_image_prep | “resized_crop_512” |
| viz_freq | 50 |
Author: Lorenzo Cardarelli, Elisa Pizzuti
Benchmark
