<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/df744237-3275-4297-a66d-8ed0ec0f0afb/original=true/df744237-3275-4297-a66d-8ed0ec0f0afb.jpeg" /><h2 id="waisnackmix-krea">WaiSnackMix Krea</h2>What went in to WaiSnackMix Krea: Krea 2 Raw + My proprietary WSM Krea LoRA + Refusal Reduction &amp; Turbo 8 Step.The main ingredient of course, unique, with extra refined training data that has not be used anywhere else, is the WSM style. In this first release, the style will consistently produce Anime, including for most backgrounds and things.<h3 id="generation-guide">Generation Guide</h3>Steps: 8 Sampler/Scheduler: Euler A / Simple, er sde / simple, res_multistep / simple, lcm / beta if you're feeling a bit quirkyCFG: 1A note on steps: This quantization has the 8 step turbo applied. If you write a complex prompt involving complex objects like musical instruments, and they come out malformed, you might try more steps. Likely, the model rushing to lay things out during early steps and then being unable to correct later on, so 10 steps might help!Keywords:There are no keywords to activate the WaiSnackMix style, but here are some I tend to use just because, and they were in the LoRA training data.<pre><code>anime style, very thin black outlines</code></pre><pre><code>anime style, ultra-thin black outlines</code></pre>Here are a few things you can put in to significantly affect the output:<pre><code>retro anime style</code></pre><pre><code>ghibli style</code></pre>The style comes from Krea 2, however, the WSM model fights this style and so you end up with a hybrid. If you really want to get that style you should try a different checkpoint. The tradeoff of getting a consistent style in this case is to leave little room for flexibility.<h3 id="a-word-of-thanks">A Word of Thanks</h3>I want to thank CivitAI for hosting training services. It was imperative that I use gold buzz for training and thanks to CivitAI's buzz economy, I could do that without worrying about running out. Thanks also to Krea AI for making the Krea 2 model, which has proven to be a stand-out for ease of use and prompt adherence.Thanks to Lonecatone23 for making his ModelBuilder nodes, which made the LoRA bake, save, and Quantization a snap.Also thanks to the many artists! Could not do it without your support and engagement.Hope you all enjoy and I look forward to seeing what you make.<h3 id="waisnackmix-anima">WaiSnackMix Anima</h3>Anima has proven to be a great model, still, and for its size it really packs a punch, especially when it comes to NSFW. And that is a big part of what we do here. Today, I won my first crucible, and I did it with an image made with WSM Anima Beta -- and I was going up against everything out there, heavy hitters like Krea bf16 models. But it's okay, Anima is up to the task. So, given the opportunity to try some new merge nodes from Lonecat, I had to update WSM to 1.0 at long last.I am proud to present my latest merge, which follows the WaiSnackMix brand in that it:<ol><li>Does the Wai style with aplomb</li><li>Wonderful colors and aesthetics</li><li>Good at following prompts</li><li>Is flexible and can do semi-realism</li><li>Has no baked LoRAs</li></ol>And since it's basically just the same models in there, balanced in such a way that the Wai comes out, then it's probably not entirely unlike Beta. In fact, Beta was one of the models that went in, and the other model was <a target="_blank" rel="ugc" href="https://civarchive.com/models/2405353/strangeuporia?modelVersionId=3162447">Strange Uporia</a> by <a target="_blank" rel="ugc" href="https://civarchive.com/user/Zlax">Zlax</a>, a creative and frankly brilliant creator I've merged from in the past.This time I had the pleasure of using<a target="_blank" rel="ugc" href="https://civarchive.com/user/lonecatone23"> Lonecat's</a> Anima merge custom node, from his indispensable LC ModelBuilder custom node pack (a must install for Comfy users for the coming months and beyond). It was absolutely a great experience and the GUI for making the individual levels on a line graph made the experience so much easier.<h3 id="generation-guide">Generation Guide</h3>Anima Positive Prompt (recommended): <code>masterpiece, best quality, score_7</code>Anima Negative Prompt (recommended): <code>worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, chromatic aberration</code><h3 id="recommended-settings-anima:">Recommended Settings Anima:</h3><ul><li>Steps: 20-30</li><li>CFG Scale: 3.5-5</li><li>Sampler: ER_SDE / Euler a</li><li>Scheduler: Simple / Beta</li><li>Recommende Resolution: 1 megapixel, however you want to slice it. 1.2 is ok sometimes but I wouldn't push it.</li></ul><h3></h3><h3 id="about-waisnackmix-anima-beta">About WaiSnackMix Anima Beta</h3>WaiSnackMix Anima Beta takes from a couple of other models which will be detailed soon. Basically I could not use them directly but had to fold them in and make some trial checkpoints which then ultimately got merged again to make this one. This model does not have any baked LoRAs, although baked and negative LoRAs were applied to predecessor models carefully and judiciously. WSM Anima Beta is flexible, so prompting for realism can get some fairly realistic-seeming semi-real images. LoRA support can further move the style dial in either direction.<h3 id="about-to-waisnackmix-illustrious">About to WaiSnackMix Illustrious</h3>WaiSnackMix 1.1 is a superb and highly versatile Illustrious-based merge that strikes an excellent balance between the vibrant, expressive Wai style and broader artistic flexibility. It handles everything from classic Wai-style illustrations to semi-realistic renders with impressive stability, strong prompt adherence, and beautiful results across a wide range of subjects, compositions, and clothing details.I genuinely love seeing what creators are producing with it — the community output has been fantastic and continues to inspire me. While I may experiment with future tweaks down the line, version 1.1 feels like a sweet spot that preserves the core appeal of the original, and it's something the community is embracing. If you're looking for a reliable, fun, and creative checkpoint that can shift smoothly from anime illustration to more realistic aesthetics without losing its charm, give WaiSnackMix 1.1 a try! WaiSnackMix v1 is also available for download, and offers similar results.<h3 id="what-you-get-with-waisnackmix">What you get with WaiSnackMix</h3>No baked LoRAs (unless they came in through the parent and predecessors models) and lots of stability, clothing options, and versatility. I have seen this model pull off amazing results in semi-realism considering that it was made to produce Wai anime style toons.Speaking of Wai, some are curious about the name I used. This checkpoint had several inputs, making it hard to pin down a short and memorable name. So, I improvised. I will return to this point but first allow me to introduce the predecessor models of this checkpoint:<a target="_blank" rel="ugc" href="https://civarchive.com/models/2371369?modelVersionId=2569232">WaiNova 1.0</a> (base) by <a target="_blank" rel="ugc" href="https://civarchive.com/user/OGMustard">OGMustard</a> - An awesome checkpoint for real, great adherence to begin with, go check it out. The color and style of this checkpoint shines, in a short time it has become recognized as a wonderful checkpoint to create and enjoy.<a target="_blank" rel="ugc" href="https://civarchive.com/models/1976483?modelVersionId=2418502">Strange Thing Mix Toon v2</a> (added) by Zlax - a softer look than WaiNova, one I really like, and with Zlax's care and testing regarding prompt adherence a known factor made it a top choice for me.And again, no baked LoRAs.Zlax and OGMustard make some of the best checkpoints around if their style matches yours. You absolutely should follow them, and you will find their models recommended below.I was going to call it WaiStrangeNova but that’s just weird and pulls too much from Zlax’s brand. Mustard is a condiment and so Strange became Snack and merge became Mix. I couldn’t be happier with the.name.Since inception this checkpoint has earned enough buzz for me to keep it on site without coming out of pocket, so I thank you all for that. It also hit some download and like milestones which to me just kind of amazes me.I recommend Euler A all around and some others below. Of course other combinations exist, and I have found this checkpoint can work well with Exponential approaches but they take longer. Heun tends to choke on complex situations and controlnets. Feel free to experiment.Upscaling: Euler A and simple is my preference here, denoise of .3 or less in most cases.<h3 id="quality-tags-guide">Quality Tags Guide</h3>Positive Prompt (slightly adds detail/refinement/shading quality, purely optional)<pre><code>masterpiece, amazing quality, best quality, ultra-detailed,</code></pre>For greater realism, fleshy lips, significant reduction in size of anime eyes:<pre><code>masterpiece, realistic, amazing quality, best quality, ultra-detailed,</code></pre>Replace with the following for a slight adjustment to toon style<pre><code>toon style,</code></pre>Other tags that affect generation, albeit slightly:<pre><code>anime style, retro style, kawaii style</code></pre>Tags that will profoundly alter the style, if used all together:<pre><code>vibrant hues, dynamic pose, blue palette, realistic, abstract background,</code></pre>As for prompts, it is very good at understanding booru format and lengthy natural language prompts. However, very brief and short prompts can, depending on the subject matter, confuse checkpoints. I highly recommend using the preview KSampler from efficiency nodes so that you can see what is happening during the layout phase of generation as it helps to demystify things.<h3 id="generation-settings">Generation Settings </h3>EulerA, Simple or Karras - works greatCFG 5Steps: 25-35 DPM++SDE with Karras-CFG 2-6 (yes, 2 is perfectly fine, maybe better, 1 is 40% faster and comprehends prompts but sepia tones creep in)Steps: 25-35CFG at 3 or higher tends to give that bright, anime look but shading can suffer, especially if you want a scene of evening romance in the bedroom.You can also experiment with lightning generations, slightly lower quality but iteration rate increases by 200% and you need fewer, resulting in generations only taking 3-5 seconds:Euler A / Simple8 StepsCFG ~1.5I'm curious to hear what you have to say about these settings, if people insist on CFG 1, I don't see how those outputs are acceptable with this checkpoint, fast or not if it looks all faded and sepia it's a no-go.The following settings work great, phenomenal, but do not combine with the SDXL lightning lora, don't even try.exp_heun_2_x0_sdeexponential20 stepsCFG 3There are so many tags that it responds to, I can't possibly hope to begin, just grab your favorite tag reference and go for it.

WaiSnackMix Collection by MGHerder

Trained on Krea 2 Base My WSM proprietary LoRA &amp; Refusal reduction baked in No other baked LoRAs

Files with SHA256: 1f508df0601a1e342bed577f10bef3ab725014ecabadf38e7385e47c5138aac7