Bad Card Template

PLEASE READ INSTRUCTIONS (this model is not very good)

Problem

It’s not possible to generate card templates for your boardgames/collectible card games with stable diffusion. This is because cards are not square, and they have been cropped by stable diffusion, and only the middle of the card is rendered.

This model

Trained on 29 not very good images.

There are several checkpoints:

• early checkpoints that do not render a correct card, but still keep the style of the original model (called bct-render*.safetensors)

• later checkpoints that render a correct card, but replace the base model style by the style of the not very good dataset I used (called bct-cardshape*.safetensors)

• so you need to fiddle a bit …

Prompting

typical prompt used in training

• a blank trading card, light blue, art deco style, with a title bar, an illustration box shaped like a planet, a subtitle bar, and a grey text box

• a blank trading card, pale blue, texture style, with a title box on top with a symbol, a white illustration box and a small pale blue text box at the bottom

• a blank trading card, blue with gold frame, texture style, with space for a title, a black drawing space, and a cloudy blue text box with card statistics

so “a blank trading card” trigger word, the main colors (green, pale blue and gold), a style (art style or ‘heroic fantasy’, ‘steampunk’), and a card structure description with “title”, “illustration/drawing”,”subtitle” and “textbox”

typical negative prompts:

• those that avoid writing, because the base model still remembers game cards having text: text, english_text, japanese_text, kanji, markings, watermark, signature, artist name,

• those that avoid having several cards: several cards, many cards, two cards

Using this model

• Download all the Loras that begin with v1.0 and place them in the Lora folder (3 cardshapers and 3 card renders)

• prompt with a ‘cardshaper’ Lora until you find a card with a shape you like

• now use this card outline as a controlnet (basically, we’re taking the drawn card and using the outline, but will now use a Lora that cannot draw cards very well but transfers style well)

  • Controlnet X Enable

  • put the card shape in the controlnet

  • Preprocessor: softedge_hed

  • Processor: controlnet_hed

  • Starting Control Step: 0.0

  • Ending Control Step: ~0.35 (higher: shape is kept more, lower: the base model can be more creative and add some flourishes)

  • Generate with a ‘cardrender’ Lora (Lora weight 0.9 works well, Lora weight 0.5 gives even stronger style transfer from the base model)

• you should have a nicely generated card

• you can play with Ending Control Step and cardrender Lora weight as indicated

• if you can draw a bit, you can take the generated black/white card outline (softedge_hed output), edit the outline yourself, put it as controlnet input and disable the preprocessor

• On the plus side, using a controlnet means you can make a card for each color/faction with the same outline but a slightly different style/color

Apologies

Yeah, this model is not very good, but hopefully someone can generate enough cards with it to make a good Lora …