DS-Fusion
PIPELINE
Setting up basic environment
Use environment.yaml from the official Stable Diffusion project, to set up the environment.
conda env create -f environment.yml
You will also need to download the checkpoint named "model.ckpt" from Source. For ease, we provide an alternative link Link for the checkpoint file we used in our project from the latent-diffusion official source.
Basic generation
Please run the following script to finetune for a specific style and text. Only single alpha-numeric characters can be accepted. For ease of use, some font data has been generated for quick testing. List of these fonts is in ldm/data/list_fonts.py. If you use the name of one of these fonts, please use them with the --one_font argument. Only capital letters and numbers can be used for this purpose, as only they are available pre generated.
python script_basic.py -s "DRAGON" -t "R" --one_font "False" --font_name "ani" --white_bg "True" --cartoon "True" --ckpt_path "ckpt/model.ckpt"
python txt2img.py --ddim_eta 1.0 --n_samples 6 --n_iter 1 --ddim_steps 50 --scale 5.0 --H 256 --W 256 --outdir out --ckpt logs/DRAGON-R/checkpoints/last.ckpt --prompt "DRAGON R"
- Use command "--make_data True" in finetuning step to override previous generated style images.
- Set --one_font as False, if wanting to use multiple fonts for use in generation. In this case it would be better to increase max_steps in config to 1000+.
- Add additional style attributes using --attribute in finetuning command. ensure to use the same attributes when generating
Custom Font
- You may use --custom_font and give a name of a font available on your system. In this case you may use any alpha numeric character, provided your system can generate it.
- You may need to adjust parameters of rasterizing in ldm/data/rasterizer.py because depending on the font, it may not turn out as expected. Look at img_base.png to see what the font looks like rasterized.
- If using --custom_font, add full name including extension. e.g. " --custom_font 'TlwgTypist-Bold.ttf' "
python script_basic.py -s "DRAGON" -t "R" --custom_font "TlwgTypist-Bold.ttf" --white_bg "True" --cartoon "True" --ckpt_path "ckpt/model.ckpt"
python txt2img.py --ddim_eta 1.0 --n_samples 6 --n_iter 1 --ddim_steps 50 --scale 5.0 --H 256 --W 256 --outdir out --ckpt logs/DRAGON-R/checkpoints/last.ckpt --prompt "DRAGON R"
Generating results using pre-trained Generic Use model
A pre-trained model has been trained over all capital letters and numbers, to provide a fast generation. This method was trained using 40 categories (in classes.txt) but has generalized sufficiently to out of training examples. Please download the checkpoint file from Link and place it in ckpt folder. Write prompt as "style style-attributes letter" Please make sure the letter is either a capital letter between A-Z or a number 0-9, otherwise it is unlikely to work well.
python txt2img.py --use_generic "True" --ddim_eta 1.0 --n_samples 6 --n_iter 1 --ddim_steps 50 --scale 5.0 --H 256 --W 256 --outdir out_generic --ckpt ckpt/ds-fusion-generic.ckpt --prompt "DRAGON R"
Examples of Results
Acknowledgement
The implementation is based on Stable Diffusion/Latent Diffusion Git-Source. The discriminator structure is referenced from DC-GAN.