Eleet model is a block-weighted merged stable diffusion model aiming at generating good quality 2D anime style images that satisfy my personal taste and, hopefully, your taste.The model name ELEET commemorates the wide use of its numerical spelling (31337) to set the Eta Noise Seed Delta (ENSD) parameter when using stable diffusion models.About samples<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/1203f19c-fc19-4cd2-97d5-530e4057d23b/width=525/1203f19c-fc19-4cd2-97d5-530e4057d23b.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/5aa08978-27ab-45be-b84e-e0f5fbac3941/width=525/5aa08978-27ab-45be-b84e-e0f5fbac3941.jpeg" />During my latest test, I found ClipSkip=1 can provide possibly better images quality than ClipSkip=2. Try it out! See my other posts for more samples of ClipSkip 1:<ul><li>Anime girl samples: <a target="_blank" rel="ugc" href="https://civarchive.com/posts/386735">Eleet v2 - 16:9 samples with ClipSkip=1</a></li><li>Scenery samples: <a target="_blank" rel="ugc" href="https://civarchive.com/posts/395993">Eleet v2 - 16:9 scenery samples</a></li></ul>Notes: I don't know why, but Civitai automatically appends these samples to the Eleet model showcases. These samples are generated with high resolution (roughly 1080p) and I also used LoRA. These samples are generated with ClipSkip=1.<hr />The model is tested on anime girls and scenery outputs. Almost all samples are made of pure txt2img or txt2img+highres fix, while one or two samples may be generated with an style-transfer extension called <a target="_blank" rel="ugc" href="https://github.com/mcmonkeyprojects/sd-dynamic-thresholding">Dynamic Thresholding</a>. No controlnet are used for generating the samples. No any LoRA or embeddings. No any regional control extensions.Due to my GPU limitation, I couldn't generate ultra-high resolution images. Therefore, none of my sample images of this model are of high resolution (at most ~1000 pixels on the long side). But these images are still able to show the model style and basic prompt use cases, so it's probably enough… I guess?Suggested settingsFor users who know little about stable diffusion settings, I recommend:<ul><li>Prompt: Start with <code>masterpiece, best quality</code>and optionally <code>aesthetic</code>. Personally I also like to append photography phrases to the end of prompt, such as <code>cinematic lighting</code>, <code>professional shadow</code> , etc.</li><li>Negative prompt:<ul><li>A basic example for negative prompt:<pre><code>(worst quality, low quality:1.4), lowres, bad anatomy, (blurry), (text, logo, watermark, signature, username)</code></pre></li><li>Black bar/black frame: To avoid black bars at image boundary, add (letterboxed, pillarboxed) to the negative prompt. This is usually needed on low-end GPU under --lowvram WebUI mode for middle-large size landscape-style images (for example, 960x544).</li><li>Depth of field (DoF): If you prefer a style similar with large depth of field, which means "being sharp &amp; clear everywhere from near to far", you may also add (depth of field, bokeh) to negative prompt. However, real-world portrait photos often have small DoF.</li><li>Most negative embeddings (for example, EasyNegative) will change the quality and the composition of the outputs. I don't use any embedding for either daily use or testing the model, so use them at your own risk.</li></ul></li><li>Sampler: DPM++ 2M Karras for stable and fast generation. DPM++ 2S a Karras is slower, but in my opinion with a higher performance and creativity ceiling.<ul><li>I feel SDE samplers (for example, DPM++ 2M SDE Karras) are sometimes overcooking images, so I personally still prefer DPM++ 2M/2S a Karras.</li></ul></li><li>CFG Scale: 6~9</li><li>Steps: 16~30</li><li>Highres Fix (optional): Latent sampler; 0.55~0.7 denoise strength; 16~25 Highres steps. When use low denoise, increase the Highres steps accordingly.</li><li>Clip skip: 1 or 2. For ClipSkip=1 samples, see <a target="_blank" rel="ugc" href="https://civarchive.com/posts/386735">Eleet v2 - 16:9 samples with ClipSkip=1</a> and <a target="_blank" rel="ugc" href="https://civarchive.com/posts/395993">Eleet v2 - 16:9 scenery samples</a>.</li><li>LoRA (optional): Suggest using <a target="_blank" rel="ugc" href="https://civarchive.com/models/15729/colors">Vibrance (v1)</a> LoRA with weight 0.15~0.20.</li><li>No external VAE file required</li></ul>Change logs<ul><li>2023/07/12: Add ClipSkip=1 samples. Add negative prompt suggestions to avoid black bars.</li><li>2023/07/08 [v2.0 release]: Modify negative prompt suggestions. Add Highres Fix suggestions.</li><li>2023/05/29: (1) Update sample images using the latest WebUI version (v1.3) and its new DPM++ 2M SDE Karras sampler; (2) Replace the previous model cover image with a portrait one, as the site doesn't seem to be friendly for displaying landscape cover images.</li></ul>

Eleet Model

Compared to v1.x versions, this new version v2 brings following good news:<ul><li>Model style more close to typical 2D. This is something about shadow. We know 2D images also need shadow, and what I have tried for this version is to make shadow less volumetric (less like 3D).</li><li>Improved prompt understanding. Previous version sometimes ignores or misdraw colors, such as drawing cyan/purple as black (however, please note that Highres Fix can cause color shift). It is better on this issue now. Also, this version also seems to understand or do better on more Booru tags.</li><li>Probably better "logic". In previous versions, there were often illogical parts in the image (such as the sea level on either side of the characters not being equally high, or very serious non-parallelism when drawing railroad tracks). Although I do NOT think merging can eliminate the problem, I believe the problem is more or less mitigated when I test the new model.</li><li>Possibly better quality on low-end or outdated GPUs. This time I made more tests on my ancient laptop (with a GPU from the GTX 900M era) under the WebUI's lowvram mode.</li></ul>Eleet v2 also brings some impacts questionable, or possiblely negative:<ul><li>Cloud style changes. I believe that it tends to draw clouds in the sky in a different way from the previous version.</li><li>Sunlight style changes. Seems that the new version prefers to draw golden sunlight in the sky even there is no sun in the image. Sometimes this causes weird light and shadow relations or unsuitable yellowish-purple color in the sky.</li></ul>

Files with SHA256: 694d38de47c17aebd7a966d674cca4da0e75c2154b1be2e84d20dbb874e92c18