You can deploy this model and start earning money today!

Deopusi/
sd-satlas-inpainter

Name: sd-satlas-inpainter
Availability: InStock

safetensors

Deopusi/satlas-tci-mini

Views

No views yet

Running Stable Diffusion for Inpainting on SatlasPretrain data

To run, get data from Deopusi/satlas-tci-mini.

Training

Cmd args:

bash

python train_dreambooth_inpaint.py \
    # Init args
    --pretrained_model_name_or_path <Base model to finetune, default "runwayml/">
    --instance_prompt <Prompt to use for all images, default "a high-resolution sks satellite image of a landscape, taken from space">
    --prompt_file <CSV file containing prompts tied to training images, defaults to None, only uses instance prompt>
    --prompt_key <Name of the prompt column in the prompt file. Only used if a prompt file is given>
    # LoRA args
    --use_lora <Optional, whether to use LoRA or finetune SD weights>
    --lora_r <int> 
    --lora_alpha <int> 
    --lora_dropout <float> 
    --lora_bias <str> <LoRA UNET args>
    --lora_text_encoder_r <int> 
    --lora_text_encoder_alpha <int> 
    --lora_text_encoder_dropout <float> 
    --lora_text_encoder_bias <str> <LoRA text encoder args>
    # Saving args
    --output_dir <Output dir for saved models & checkpoints, default "model">
    --checkpointing_steps <Number of steps between checkpointing, default 500. To preserve memory, only one checkpoint is allowed.>
    --checkpoints_total_limit <Number of checkpoints that will be fully saved (/!\ ~3.7GB per ckpt), defaults to 1>
    --resume_from_checkpoint <Name of checkpoint to resume from, default None (start from scratch). Use 'latest' to default to most recent ckpt.>
    # Hyperparameters
    --resolution <Image resolution. Be aware that data is of resolution 512. Default 512>
    --train_text_encoder <Optional, whether to train text encoder in addition to UNET>
    --train_batch_size <Batch size for training, default 4>
    --num_train_epochs <Number of training epochs>
    --learning_rate <Learning rate>
        

To start training from the beginning, run:

bash

python train_dreambooth_inpaint.py
    --use_lora
    --train_text_encoder
    --checkpointing_steps 100
    --num_train_epochs 3

3 training epochs is the limit within 4h. This trains a model only using the standard instance prompt. To add image-specific prompts, include values for the prompt_file argument.

To resume training, run:

bash

python train_dreambooth_inpaint.py
    --use_lora
    --train_text_encoder
    --checkpointing_steps 100
    --num_train_epochs 3
    --resume_from_checkpoint checkpoint-last

For image-specific prompts, same as above.

Testing

For visualizing results on a single image, run:

python test_dreambooth_inpaint.py

Generates a number of image candidates for a certain image with a fixed mask. Parameters are mostly fixed. To change things like the number of images, number of inference steps, mask shape, which image to use..., modify the script.

To measure the evaluation metric, specify --num_test_samples > 1. This will compute average feature loss, as well as average min and max loss (min/max over candidates, then average over samples).

Same as before, you can specify prompts.

Gradio App

Run

bash

python inpaint_app.py
    --num_images_per_prompt 10
    --num_inference_steps 30

You can open the app using the links, both local and online.

Panning is not yet implemented in Gradio Image objects. To use panning, first click on the image. This sets the starting position of the panning. This is final. All future panning assumes you start from there.

To pan, click on another part of the image. The image wil shift according to the difference between the set start position and the place you click.

Click the 'Inpaint' button to fill the image. You can influence the inpainting using the prompt.

Deopusi/sd-satlas-inpainter

Running Stable Diffusion for Inpainting on SatlasPretrain data

Training

Testing

Gradio App

Deopusi/
sd-satlas-inpainter