What is this?
Hand-drawn video game style that will add strong, black outlines to the objects in front and blur objects in the background to give depth to the image.
You can work with this style in the same fashion as described in this tutorial.
The style is so strong when used with 1.0 strength that the choice of the model is only a matter of composition and responsiveness to prompting.
The best results from the 8 models I tested with come from RealisticVision, Deliberate, CheeseDaddy'sLandscapeMixOffset.
Use a few of your favorite negative embeddings and general negative tags.
Quick example
positive prompt
(Award-winning photograph), (Japanese garden), (Autumn foliage), (Vibrant colors), (Fallen leaves), (Nature's beauty), (Red and gold hues), (Maple trees), (Reflections on water), (Peaceful tranquility), (Traditional wooden architecture), (Rustic bridges), (Soothing natural sounds), (Seasonal delight), (Serene garden views), (Low-angle perspective), (natural lighting), Nikon F2 Camera, hidden objects games, video game concept art, (8K Unity wallpaper), fine details, award-winning image, highly detailed, 16k, cinematic perspective, ((video game environment concept art style)), pretty colors, cinematic environment, (handmade2d:1.0) stylenegative prompt
ng_deepnegative_v1_75t, bad-artist, bad-hands-5, (worst quality, low quality:1.4), multiple views, multiple panels, blurry, watermark, letterbox, text, character, human, people, 1girl, 1boy, NSFW, nude, signature, medium quality, deleted, lowres, comic, cleavage, frame, watermark, signature,watermark,signaturesettings
CFG 7 and 50 steps on any model or the ones listed above
Description
FAQ
Comments (6)
can you share your training configuration for this style ? thank you
Sure, here it is:
{
"datetime": "2023-03-18 23:17:16",
"model_name": "deliberate_v2",
"model_hash": "9aba26abdf",
"num_of_dataset_images": 31,
"num_vectors_per_token": 2,
"embedding_name": "handmade2d",
"learn_rate": "0.05:100, 0.02:200, 0.01:600, 0.005:2000, 0.002:5000, 0.001:10000, 0.0005",
"batch_size": 1,
"gradient_step": 1,
"data_root": "D:\\Stable Diffusion Files\\training data\\games\\1\\output1",
"log_directory": "textual_inversion\\2023-03-18\\handmade2d",
"training_width": 512,
"training_height": 512,
"steps": 40000,
"clip_grad_mode": "disabled",
"clip_grad_value": "0.1",
"latent_sampling_method": "deterministic",
"create_image_every": 500,
"save_embedding_every": 500,
"save_image_with_stored_embedding": true,
"preview_prompt": "(handmade2d:1.0) style, (Award-winning photograph), (Japanese garden), (Autumn foliage), (Vibrant colors), (Fallen leaves), (Nature's beauty), (Red and gold hues), (Maple trees), (Reflections on water), (Peaceful tranquility), (Traditional wooden architecture), (Rustic bridges), (Soothing natural sounds), (Seasonal delight), (Serene garden views), (Low-angle perspective), (natural lighting), (Wide-angle lens capturing scenery), hidden objects games, video game concept art, (8K Unity wallpaper), fine details, award-winning image, highly detailed, 16k, cinematic perspective, ((video game environment concept art style)), pretty colors, cinematic environment",
"preview_negative_prompt": "ng_deepnegative_v1_75t, bad-artist, bad-hands-5, (worst quality, low quality:1.4), multiple views, multiple panels, blurry, watermark, letterbox, text, character, human, people, 1girl, 1boy, NSFW, nude, signature, medium quality, deleted, lowres, comic",
"preview_steps": 50,
"preview_sampler_index": 15,
"preview_cfg_scale": 7,
"preview_seed": -1.0,
"preview_width": 768,
"preview_height": 512,
"template_file": "E:\\stable-diffusion-webui-master\\textual_inversion_templates\\style_filewords.txt",
"initial_step": 0
}
@Peaksel thanks a lot for sharing, looks a lot my configuration but not getting this great results on the style I am working on even with very high quality images.
Did you use the Automatic1111 webui TI tool for training?
did you end up using the last checkpoint or some earlier one?
thanks again
@segalinc everything standard, web UI training on SD 1.5 model. It is all about the images you put in. Focus on your selection and see which images you need to exclude.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.