<h1 id="update-2025-12-28-gay621-xl">Update 2025-12-28 - Gay621 XL -</h1> I'm a liar. I said the last two were the final ones and that I wouldn't be doing any future versions but here we are. Please read before using the XL version if you are familiar with the older SD1.5 versions as some things have changed.This isn't intended to compete with major models, and lacks the polish of the big players, and is harder to get good results with, however it offers a unique twist in specific situations you won't get from them either. It exists only because I could not reliably get what I wanted from other models, so to that end, while I will likely refine this, the cadence will be very slow, so go into it with tempered expectations. Note: Loras for other models are unlikely to perform well with this. If you really want to try some, IllustiousXL / NovaFurryxl will be your best bet at having some success with. Prompts are all e621 tags, with all '/' characters and spaces replaced with '_' characters.Artist tags are all prefixed with 'by_'.Split tags by commas. IE: duo, male_male, anthro, athletic, sheath, canine, 5_fingers, by_zackary911There are two meta-tags you can experiment with, but using them does NOT guarantee better results, it's just another thing you can mess around with which is meta_liked and meta_loved.meta_like was tagged on a select subset of highly liked e621 images.meta_loved was tagged on an even more restricted set of highly liked e621 images.Still figuring out ideal range, but above CFG 9 starts to get pretty crispy, so likely CFG between 4 and 9. By default, this will NOT produce as nice of images as you will find in NovaFurryxl.The benefit to using this over pony or novafurry is that is has pretty strong tag adherence, and a strong male bias.- The downside of this is you need to be very specific. You should typically specify an artist style, you should specify a background, etc. If you want a background but aren't picky, try detailed_background or amazing_background.- I strongly recommend always prompting with number (IE: Solo, Duo, trio, etc.), species (IE: Wolf), and a generic background tag (IE: amazing_background) and then refining and adding from there. Otherwise you'll usually get average at best results on a gradiant background.- If fingers are an issue, consider specifying number of fingers IE: 5_fingers- Even though this was heavily refined, this model has roots in base sdxl, novafurry, and one I trained off pony, it's a blend and refinement of several models, so you can usually get away without using underscores for concepts that don't have e621 tags by using normal spaced text and descriptions, give it a try.This model was trained using art of women characters, however it was given a very heavy male / gay male bias.Note: I wasn't able to completely smooth out everything after trying to over-write the concept of a 'sheath', so if it starts putting them on everything during hi-res operations, try dropping the denoising strength or upscale by a bit. I find between 0.55 and 0.65 is usually pretty safe when upscale by 2 is set.Note: Was trained with clip_skip=1, but you can try 2 if you want. There are still some remnants of the underlying sdxl and various other models used as the foundation. Therefore, some plain text prompts should work as well. For example, a common thing people prompt with is like 'masterpiece,best quality,ultra-HD,photorealistic,high detail'. These are NOT necessary, however they do still have subtle impact.Note about underscores: Technically, the underscores aren't needed even though I usually use them. Typically the underscores is broken out and interpreted as it's own thing in the sequence. IE: Anthro_Male is usually interpreted as three tokens, 'Anthro', '_', 'Male'. So using 'Anthro Male' will still match two of those 3. The benefit in using them is that you'll likely get a slightly stronger match hitting on the sequence of the 3 instead of just the two main words.Consider it like talking quietly, when prompting with e621 tags is like speaking normally. This means that you can actually get some pretty good results from mixed format prompting. View showcase images for examples of how to prompt with this model. The models base resolution is 1024x1024, and it's bucketed by 64 pixels. So buckets would be:DATASET BUCKET DISTRIBUTIONTotal Raw Images (Weighted): 139,152Total Refining Images (Weighted): 84,690==========================================================ID | Resolution | Shape Description | Percent-----|---------------|-----------------------|---------0 | 768 x 1280 | Tall Vertical | 6.45%1 | 832 x 1216 | Vertical | 14.98%2 | 896 x 1152 | Vertical (Portrait) | 16.44%3 | 960 x 1088 | Slightly Vertical | 8.72%-----|---------------|-----------------------|---------4 | 1024 x 1024 | Square | 11.53%-----|---------------|-----------------------|---------5 | 1088 x 960 | Slightly Horizontal | 7.99%6 | 1152 x 896 | Landscape | 12.16%7 | 1216 x 832 | Horizontal | 11.15%8 | 1280 x 768 | Wide Horizontal | 10.58%==========================================================Suppressed concepts: While not impossible to generate, almost none of the refinement data I used include the following, due to complications they were causing: taur, multiple genitals, or drawings with extra limbs. These concepts will be very hard to generate with this.<h1 id="update:-2024-12-01-hybrid-model"> Update: 2024-12-01 - Hybrid Model -</h1> I know I mentioned that the other version was the final, but there was one other project I was messing around with a few months ago that I find I've been using more than the others. The better models I mentioned when retiring this project was the Indigo Furry Mix project. My one complaint with it was that it was was very 'same-y' IE: Every sheath looked the same, every expression was similar, etc. In an attempt to fix that created a hybrid model between one of the indigo furry mixes (I believe 105) and the multiple character model posted here (Roughly 40% indigo, 60% multiple character, and then trained for a few epoch on a small subset of images from the solo version, as well as a small subset of masculine females to better work with femboys). I think this results in the best choice of the models on this project if you just want a general purpose 'create gay furry stuff' model out of the box. However there are still models that do everything this does better out there. I do know many of you still use this project though and I'd feel bad if I was withholding something from y'all for the sake of assuming nobody wants it. Recommend generating images between 640x640 and 768x768, still using e621 tags. This model is slightly prone to burnout however, so be careful. Given the 40% weighting of the indigo furry mix project, some of the resources they provide are relevant to this version. A Great resource for picking an artist style is provided by their project here <a target="_blank" rel="ugc" href="https://mega.nz/folder/FoQR1RgA#gDLJyJe9-JK2_FDaosgSzg/folder/osYgSBCb">https://mega.nz/folder/FoQR1RgA#gDLJyJe9-JK2_FDaosgSzg/folder/osYgSBCb</a> Check out the indigo furry mix project here as well <a target="_blank" rel="ugc" href="https://civarchive.com/models/34469?modelVersionId=480978">https://civarchive.com/models/34469?modelVersionId=480978</a> With the exception of what I mentioned above, it's (indigo) generally a better model and a little less prone to burnout. Being blended with indigo as well also greatly improves this models ability to break from super strict E621 tags (though it's still recommended) I currently am utilizing my hardware for other AI projects, so continued development on this is very unlikely, if it were to continue it would be experimenting with newer models and likely starting from scratch with a more current image dataset. Update: 2023-08-07 In my mind there are now other models that do what this one does better, so I've stopped development on it. The 'Solo' Version is the original model, which does solo males quite well. The 'Multiple Character' model was a separate project that I was working on, but am just not completely satisfied with the results so won't be continuing work on it. It can also do solo males quite well however. Thanks for enjoying the project for the last few months! While I'm not currently planning a sd 2.1 flavor, I may look into creating a similar model using future versions. As with the original model, prompting on the multiple character model uses e-621 tags and you should be very specific. I was testing using Euler A 32 832x448 cfg 10 The primary goal of the 'multiple characters' model was to accomplish two things: 1. Get reasonably acceptable base images without having to use additional addins or hi-res. 2. Make generating huffy sheaths (uwu) easier than in the solo model, which struggled a bit to get them correct.3. Make it understand that it's okay for PPs and Balls to touch without merging together <h1 id="original"> --Original--</h1> This is intended to be a semi specialized solo gay male pinup / character creator. If you are looking for a model strong at generating females, this is not it lol. It is able to create multiple character images, but that's not what it's ideal for. There is currently another project I'm working on which will be better suited for that. This models ideal range is femboy to twunk. See testing prompt for a good place to start. Early version, still lots of work to do, but does not re-use any previous furry or yiffy models, which always messed up the pps (uwu). Trained various combinations of around 120,000 male gay focused furry images, for use with stable diffusion web UI. Tagging uses comma separated e-621 options without any underscores. Exception to tags: YMMV, but underscores are still needed when specifying a specific character tags and MetaTags. Meta tags are preserved as well, and continue to use underscores. IE: Vincent_(FoxMcC) on e621 would be represented as Vincent_\(FoxMcC\) in this model. IE: hi_res and digital_media_(artwork) on e621 would be hi_res, digital_media_\(artwork\) Artist tags will usually work okay with just the name, but you'll have the best luck prefixing them with 'by' IE: by chunie, by richard foley, etc. Be VERY SPECIFIC with what you want in an image with this, down to specifying eye color. See sample prompt. It is intended to be more of a character or pinup generator, but it's able to do well with multiple characters as well with careful sculpting. Some kinks are not represented in the model, images which were tagged with the following are not included in the image training dataset -animated -humanonly -hyper -watersports -scat -cub -comic -diaper -pixel(artwork) -3d_(artwork) -vore -chibi This isn't because I have anything against those kinks, rather I'm trying to create a slight built-in bias or remove unexpected elements from the images it generates. The criteria to be included in the training was a favcount greater than 200 and a score over 100, which also contain the tags 'anthro' and 'male' Note: Specifying the number of fingers you want helps immensely, and put your most important criteria at the front of the prompt, and nice to haves near the back. Testing configuration /prompts Euler A - 28 (Heun 30 usually is an interesting vatiation) CFG 8 (works well between 6 and 9 depending on the prompt) Seed - 855213688 (you can set it to whatever but this is what I test with) Optional Hires testing configuration Upscaler - Latent (nearest) Denoising strength 0.6 Upscale 2 Positive Prompt: wolf, wolfdog, striped hyena, anthro, male, detailed, (detailed fur), (athletic), athletic anthro, muscular, muscular anthro, muscular male, fluffy, fluffy fur, fluffy ears, fluffy hair, fluffy chest, mane, pinup, canine ears, canine penis, canine anatomy, chest tuft, (balls), big balls, nude, nude male, solo, (fully sheathed), (penis tip), thigh up, (yellow eyes), ((sheath)), (white sheath), (black sheath), (brown sheath), tan sheath, gray sheath, retracted sheath, smug, grin, horny, lustful gaze, [orange fur], [red fur], [white fur], [blue fur], [yellow fur], two tone fur, two tone body, two tone hair, multi tone hair, sitting, posing, reclining, detailed background, five fingers, rating:explicit, digital_media_\(artwork\) Negative Prompt: deformed, blurry, multi limb, taur, multiple limbs, bad anatomy, disfigured, poorly drawn face, mutated, extra limb, ugly, woman, female, multiple characters, feral, deformed hands, face out of frame, multiple tails, over saturated, hdr, high dynamic range, sketch, comic, human penis, no pupils, simple background, multiple characters, cub, child, jock, underwear, clothing, chastity, cage, abstract, chastity cage Possible future testing versions can be found in Furry Diffusion discord or at <a target="_blank" rel="ugc" href="https://wolf.gay/">https://wolf.gay/</a> No commercial for profit use of any kind, or selling images made by this model, or any derivations of this model / images created by them.

Gay621 - Furry Male Focus

I know I mentioned that the other version was the final, but there was one other project I was messing around with a few months ago that I find I've been using more than the others. The better models I mentioned when retiring this project was the Indigo Furry Mix project. My one complaint with it was that it was was very 'same-y' IE: Every sheath looked the same, every expression was similar. In an attempt to fix that created a hybrid model between one of the indigo furry mixes (I believe 105) and the multiple character model posted here (Roughly 40% indigo, 60% multiple character, and then trained for a few epoch on a small subset of images from the solo version, as well as a small subset of masculine females to better work with femboys). I think this results in the best choice of the models on this project if you just want a general purpose 'create gay furry stuff' model out of the box. However there are still models that do everything this does better out there. I do know many of you still use this project though and I'd feel bad if I was withholding something from y'all for the sake of assuming nobody wants it. Recommend generating images between 640x640 and 768x768. Given the 40% weighting of the indigo furry mix project, some of the resources they provide are relevant to this version. A Great resource for picking an artist style is provided by their project here <a target="_blank" rel="ugc" href="https://mega.nz/folder/FoQR1RgA#gDLJyJe9-JK2_FDaosgSzg/folder/osYgSBCb">https://mega.nz/folder/FoQR1RgA#gDLJyJe9-JK2_FDaosgSzg/folder/osYgSBCb</a> Check out the indigo furry mix project here as well <a target="_blank" rel="ugc" href="https://civitai.com/models/34469?modelVersionId=480978">https://civitai.com/models/34469?modelVersionId=480978</a> With the exception of what I mentioned above, it's generally a better model and a little less prone to burnout.

Files with SHA256: 2252b91ea18851c319cc0603284476f1a8478cfd027e23434d07dd55175120b4