IF YOU WANT TO USE v1.0, YOU NEED TO READ THE INSTRUCTIONS (<a target="_blank" rel="ugc" href="https://civarchive.com/articles/671">https://civarchive.com/articles/671</a>)This model makes card templates for board games / trading card games.<h2 id="loras-wp2cebejk">Loras</h2><ul><li>bct2_basic: makes cards, slightly overcooked (too detailed, lots of 3d, slightly washed off colours, not bad looking)</li><li>bct2_2080_basic: makes cards, removed the overcooked checkpoints</li></ul><h2 id="prompting-pz85cv0e0">Prompting</h2><ul><li>the activation word 'bct2'</li><li>a color: 'green color', 'lava red colour', 'black and gold color'</li><li>a style, artistic or universe: 'watercolor style', 'steampunk oil painting style','digital art style'</li><li>a description of the background: 'black background','forest in the background','background with scribbles'</li><li>a description of the card: 'card with gold ornaments','card that looks like a jewel'</li><li>a description of the title banners: 'blue title banner', 'hammered gold text banner'</li><li>a description of the illustration: 'oval illustration space with thunder','black picture frame'</li><li>a description of the text box: 'text box with cloud texture'</li><li>a description of the spaces to put stats: 'heart symbol at the top left', 'shield symbol at the bottom right'</li><li>quality: 'masterpiece, best quality'</li><li>negatives: blurry, badly drawn, ugly, low quality (from the training set)</li></ul>obviously this is not all necessary or taken into account in the image generation, but it's what it was trained with ...<h2 id="weight-mozqlynhm">Weight</h2>This is a difficult one, because most models really want to draw a human being or waifu.<ul><li>1.5~3.0 for models that can do landscapes or objects</li><li>2.5~4.5 for Lyriel and some anime models (to override their tendency to draw an anime girl)</li><li><s>0.7~1.3 for waifu basic</s></li><li>if you get a drawing and not a card then weight is too low, if card with details weight is ok, card with no details weight is too high, card with black hole in the middle weight is far too high</li><li>you can add 'plain card style' in the prompt to help get a standard card (rectangle card, title, illustration, title, text)</li></ul>in all cases, if you use a new checkpoint, it is highly recommended to do a Prompt X/Y/Z with the Lora weight<h2 id="making-cards-with-anime-models-ouviea7lx">Making cards with anime models</h2>This model was trained on OpenJourneyV4, a realistic model.<ul><li>it is more difficult to make cards with anime models, because it is hard to create the shape of a card (would require lora weight &gt; 4.5, but the Lora breaks before that)</li><li>the main factor that determines having a decent card seems to be the random seed not the weight of the Lora (need a seed that has high contrast in it)</li><li>things to try:<ul><li>try many seeds and reuse the good ones</li><li>tweak the Lora weight after having a good seed number</li><li>try adding 'plain card style' as mentioned above</li><li>try reducing the weight of the activation word 'bct2', since these models use Booru prompts and not plain language prompts, the lora on the text layer is maybe useless</li><li>another possibility is to make a card with Lyriel or XenoEngine, and run it through img2img with loopback on the anime checkpoint (example parameters: 0.6 denoising, 50 steps, 6 loopback iterations, 0.2 loopback final denoising)</li></ul></li></ul><h2 id="additional-remarks-4335t8t1u">Additional remarks</h2><ul><li>it is highly recommended to have the add_details lora (to make a card more cartoonish or more realistic) <a target="_blank" rel="ugc" href="https://civarchive.com/models/82098/add-more-details-detail-enhancer-tweaker-lora">https://civarchive.com/models/82098/add-more-details-detail-enhancer-tweaker-lora</a></li><li>you can use <a target="_blank" rel="ugc" href="http://cleanup.pictures">http://cleanup.pictures</a> to cleanup unnecessary details</li><li>you can use extensions based on segment-anything to select relevant parts of the card for post processing</li><li>the models generate dark images sometimes (because it was trained on pictures with black background and black illustration box, this can be removed with the prompt)</li><li>the trick of using a controlnet to get a specific card shape still works (see the instructions for v1.0)</li><li>the dataset used for training is available. If I had to re-train from scratch I would use a helper word as well (maybe 'tablet' or 'smartphone' or 'trading card')</li><li>I tried fitting a textual inversion, it was a failure (it didn't interact well with the Lora), but it gave a clue: a useful helper prompt is '📅', the calendar emoji. I tried adding several 📅 to the prompt, it makes the generation follow the prompt more, forces more icons (good for showing stats) and makes more subdivisions on the card</li><li>see <a target="_blank" rel="ugc" href="https://civarchive.com/models/95424/pokeffects">https://civarchive.com/models/95424/pokeffects</a> for even more Loras</li><li>a word on aesthetic score<ul><li>from what I've seen, 4.5 is a normal card, 5.5~6.0 is a very good looking card, 6.2+ cannot be a card and will be a drawing instead</li><li>therefore, not having 'masterpiece, best quality' in the prompt can help with card shape</li><li>the model was trained on realistic checkpoints, it can get 5.5~6.0+ score on realistic checkpoints, but only around 5.0~5.5 on anime models (would need to build a dataset and train a specific lora)</li></ul></li><li>a word on placement: attentional networks are used in modern neural nets (this gives translational invariance, the image is not changed a lot by moving it up/down one pixel), but it also means that placing certain elements in certain part of the image is difficult. To make cards and board games with precise placement, either a controlnet, SDXL/SD3 or attention boosting is required (see the end of the article on controlnet colors).</li></ul><h2 id="on-youtube:-9qtslgv11">On Youtube:</h2><ul><li>this youtuber asks how board game artists can still be paid in the age of AI/SD /watch?v=SdgzH2OSZDY . This Lora was made for the artistically challenged and to have better prototypes. If you have thoughts on the subject watch/comment on the video</li><li>/watch?v=XDd4u2xnRtE for important rules in making TCG cards</li><li>I saw this nice video on YT, I have a few comments on it<ul><li>the 'occlusion problem' in computer vision is seeing objects that are partially hidden behind other objects. This method removes the problem of generating the different parts of a card that mask each other by generating them one by one</li><li>the controlnet trick is similar to bct-v1.0, but textured masked latent is something new to me. By using very small dots, I'm guessing he's asking for an ultra detailed texture (fine scale details, high freq fourier)</li><li>so, the way he can generate different (and interchangeable) card parts to make different versions of cards (different clans, colors, card rarity, etc...) while maintaining compatibility because everything is given strict dimensions by a controlnet is absolutely great I think.</li><li>he blends everything together by compositing and then running a final img2img with 0.25 denoise. He's also using the same model for all the different parts of card to keep the card coherent.</li><li>for the shield, I would have preferred bct-v1.0 with controlnet with a round shape as control and low Lora weight. The bct-v1.0 Lora acts an 'image flatifier and iconifier', that could have worked better.</li><li>for the final text and image compositing, I would have preferred some known card making software (in case you need to re-generate complete game decks), but you wouldn't be able to have the final img2img blending</li><li>final gripe (to be terminally nit picky), you can make the portrait be the deepest (most masked element) of the composition, except for a few parts (in this video, the jewel and locks of hair could fall above the scroll and frame). I am not sure what this is called and how to automate it.</li><li>amazing end result<div data-youtube-video><iframe width="640" height="480" allowfullscreen="true" autoplay="false" disablekbcontrols="false" enableiframeapi="false" endtime="0" ivloadpolicy="0" loop="false" modestbranding="false" origin playlist src="https://www.youtube.com/embed/Rl4YCX8ULDw" start="0"></iframe></div></li></ul></li></ul>

Bad Card Template

Tried to generate some cards for a v2.0.It took a bit of time, and I tried to do something a bit next gen.<ul><li>generation a bunch of card and kept the ones that look like actual card templates (not paintings, etc...)</li><li>filtered by aesthetic score (kept only the cards that are 4.5+)</li><li>segmentation using segment-anything (the new segmentation model from facebook)</li><li>annotation</li><li>cleanup of generation artifacts by hand using <a target="_blank" rel="ugc" href="http://cleanup.pictures">http://cleanup.pictures</a></li><li>re-generated the aesthetic score after cleanup</li><li>created an augmentation script (flip, rotate, shuffle, add some transparency on the background)</li><li>base model: open journey v4, no VAE</li></ul>

Files with SHA256: c26db11794bbd58ee41c2050d97bd368e2d740a7664461fc794c6e15a10c8037