👋 If you like what I do and want to support the development, feel free to buy me a coffee:<a target="_blank" rel="ugc" href="https://ko-fi.com/herrscheragga"><img src="https://camo.githubusercontent.com/7765e3dbce77c82639ff14d27dda26628b206727b61a6736190f90e81c78a121/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4b6f2d2d66692d537570706f72742532304d652d4646354535423f7374796c653d666f722d7468652d6261646765266c6f676f3d6b6f2d6669266c6f676f436f6c6f723d7768697465" alt="Ko-Fi" /></a><hr /><h2 id="velvette-xl">Velvette-XL ✨</h2>Velvette-XL is the reconstruction of an unreleased experimental model—recovered, refined, and expanded into a flexible hybrid checkpoint. While earlier versions captured a very distinct identity, version 4 moves in a broader direction: it carries more visual information, more styles, and much wider compatibility.Built from a fusion of diverse 2D styles—from expressive sketch lines to polished semi-realistic rendering—Velvette-XL focuses on line versatility, adaptable detail, and creative control. It can shift between defined lineart, lineless illustration, and highly detailed compositions depending on prompt weight and CFG range.Originating from the Providence concept, Velvette-XL preserves the multi-style, lightweight philosophy of its earliest versions, now expanded into a more complete, more capable, and more flexible model—one that rewards intentional prompting while still retaining the roots of its original identity.<hr /><h3 id="v4-or-noblesse-focus">V4 | Noblesse FOCUS</h3>I've been using sd-mecha since v2, and by the time I reached v4 I ended up working with way too many versions (I even thought about giving up at one point, because it was generating nothing but noise).I was so used to SuperMerge that I had to completely rethink my entire merging workflow. I switched back and forth between add_difference and weighted_sum until I finally adapted to this new method—which turned out to be incredibly smooth and fast.Version 4 went through several full restarts (the first successful compilation is also merged here… almost as I remember it), and there are more than twenty internal iterations (5–10 models each). Some were better, some were worse, but the goal was to gather as much information as possible and then improve visual quality—although it wasn’t really necessary, since iterations 18–19 were already perfect.In short, this version contains a large number of merges. It brings together furry, hentai, 3D, 2D, anime, and more. Velvette v4 is a model designed to preserve the maximum amount of information possible. (Though I still need to add some extra pixel-level data—but I forgot.)<hr /><h3 id="better-quality-prompt:">Better Quality Prompt:</h3>Positive:<pre><code>(masterpiece, high quality, best quality, amazing quality, very-aesthetic, newest, scenery, very awa):8, (detailed, ultra-detailed, addmicrodetails, intricate details):8, (8k, Ultra HD, high resolution, absurdres, incredibly absurdres, hdr):8, (ultra detailed face, detailed eyes, extremely detailed eyes, perfect eyes, detailed iris):8, (thick lines, bold outlines, inked lines, strong contour, cel shading, digital illustration, stylized shading, soft brushwork, painterly edges):8, </code></pre>Negative:<pre><code>(worst aesthetic, flat color):8, (bad quality, low detail, worst quality):8, (bad eyes, bad iris, undetailed eyes, undetaled face, bad face):8, (bad anatomy, bad hands, extra fingers, distorted hands):8, (malformed, deformed, distorted, mutated, blurry):8, (patreon username, twitter username, artist name, signature, web address, weibo watermark, watermark, url):8,</code></pre>Style Keywords (positive or negative):<pre><code>(realistic, photorealistic, hyperrealistic, 3d, 2.5d, cgi, anime, cartoon)</code></pre><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/f24e056d-2235-4ba1-92b4-e6ef9b6fad77/width=525/f24e056d-2235-4ba1-92b4-e6ef9b6fad77.jpeg" /><hr /><h3 id="v3-or-alpha">V3 | ALPHA</h3>ALPHA is an incomplete version—a fusion of v1 + v2 intended to be merged later with IL v2.0 or v3.0, as well as rouwei v7. I’ve tested it for days and ended up liking it much more than expected, so I didn’t finish its final merge.There is also a sister version including Pony Diffusion, but I haven’t tested it enough yet. (Side note: I was a bit disappointed with Civitai, but I managed to merge checkpoints on free Colab accounts—my wallet is thankful.)<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/6e4c4b07-7b47-4a0c-a9fa-cd973ebaa3be/width=525/6e4c4b07-7b47-4a0c-a9fa-cd973ebaa3be.jpeg" /><hr /><h3 id="v2-or-eps">V2 | EPS</h3>There are changes!For V2, I tried to beef up Velvette, but the same thing happened as in v1.1 (it significantly increased the strength of the prompts), so I reduced the strength of the base style (something I wanted to do), doubled the strength of PVDC, and merged the v1.1 excerpt to 0.1 strength. This improved the model, and I simply merged it with noobai eps (I was told it was good for reducing steps, although I don't know what changed because v1 had dmd2 and I added it this time).In short... I don't know what differentiates this from that, and I've only tested it with Illustrious lora, so I don't know how much compatibility it has. Anyway, I like the result. By the way, it also works with the VVTT_detailer. Thanks for reading, bye~ 7w7)/<hr /><h3 id="behavior-with-cfg-and-steps">🎛️ Behavior with CFG and Steps</h3><ul><li>CFG 1–4 → Faithful to applied LoRA, cleaner lines.</li><li>CFG 4–7 → Velvette’s base style takes over, sketch-like look.</li><li>CFG 7–8+ → <s>More saturated and cartoonish appearance, similar to using a style LoRA at strength 1.5–2.</s></li></ul>📌 Practical tip: If you’re not using LoRAs at low CFG, add:<ul><li>Positive → <code>hdr:2.5</code></li><li>Negative → <code>flat color:2.5</code></li></ul>This adds extra detail even at CFG 1–2, reaching quality comparable to Utopia v4 at CFG 4.<hr /><h3 id="step-performance">🔹 Step Performance</h3><ul><li>Minimum useful → 6 steps at CFG 1.</li><li>Optimal → 12–30 steps.</li></ul><h3 id="quick-guide">⚙️ Quick Guide</h3><ul><li>Steps → 10–30 (optimal: 12–30)</li><li>CFG → 1.5 for smoothness | 3+ for stronger definition</li><li>Samplers → DPM++ 2M Karras, Euler a, turbo, etc.</li><li>Resolution → 1024×1024+ (1152×1344 and 1536×896 work very well)</li><li>Upscalers → R-ESRGAN 4x+ Anime6B / 4x-UltraSharp</li></ul><hr /><h3 id="useful-tags">🏷️ Useful Tags</h3>Basic styles: <code>pvdc_il, pvdc_stl, sketch style, pastel style, cartoon style, thick lines, hatched style, comic style, detailed style, fabulosity style, cutecore, horrorcore, glitchcore, cutelatte, anime, shiny</code>Examples:<ul><li><code>(sketch style + thick lines + hatched style)</code></li><li><code>(comic style + detailed style)</code></li><li><code>(pastel style + fabulosity style)</code></li></ul><hr /><h3 id="prompting-tips">🎨 Prompting Tips</h3>General positives: <code>(masterpiece), (best quality), (ultra-detailed), perfect composition, intricate details, official wallpaper</code>By theme:<ul><li>Anime/Cartoon → <code>anime, cutecore, shiny skin</code></li><li>Realistic → <code>(REALISTIC:1.15), detailed eyes, realistic lighting</code></li><li>Experimental → <code>glitchcore, watercolor, abstract, dark theme</code></li></ul>Recommended negatives: <code>bad quality, worst quality, watermark, text, bad hands, missing fingers, logo, simple shading, blurry, low contrast</code>Extras: <code>flat color:2.5, washed out:2.1, low saturation:2.5</code><hr /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/caea6a6e-abcb-4fed-a159-8886b281ca49/width=525/caea6a6e-abcb-4fed-a159-8886b281ca49.jpeg" /><hr /><h3 id="v1.1-or-deleted">V1.1 | Deleted</h3>Well, v1.1 was archived (deleted), but not before a backup was saved. The thing is, even though it was a good version, Velvette's idea was to play with the strokes. So I simply extracted a lora from the checkpoint using v1 as a base (there wasn't much difference, but this lora is very good for correcting v1 without switching to another checkpoint).<hr /><h3 id="v1-or-normal">V1 | Normal</h3>The dataset was intentionally softened to avoid overwriting model’s core concepts.🔹Some NSFW elements have been reduced, but the essential elements for experimentation have been retained. (They will be added in a future version.)<hr /><h3 id="behavior-with-cfg-and-steps">🎛️ Behavior with CFG and Steps</h3><ul><li>CFG 1–3 → Faithful to applied LoRA, cleaner lines.</li><li>CFG 4–6 → Velvette’s base style takes over, sketch-like look.</li><li>CFG 7–8+ → More saturated and cartoonish appearance, similar to using a style LoRA at strength 1.5–2.</li></ul>📌 Practical tip: If you’re not using LoRAs at low CFG, add:<ul><li>Positive → <code>hdr:2.5</code></li><li>Negative → <code>flat color:2.5</code></li></ul>This adds extra detail even at CFG 1–2, reaching quality comparable to Utopia v4 at CFG 4.<hr /><h3 id="step-performance">🔹 Step Performance</h3><ul><li>Minimum useful → 8 steps at CFG 1.</li><li>Optimal → 14–30 steps.</li><li>High CFG → experimental, depending on tags and style mixing.</li></ul><hr /><h3 id="hidden-control-with-cutelatte-negatives">💡 Hidden Control with Cutelatte Negatives</h3>A unique feature of Velvette is that it can soften or transform its base style using Cutelatte triggers in negatives. This unlocks a range of hidden styles that normally wouldn’t be available in a merged checkpoint.📊 Behavior Table with Cutelatte Negatives<h3 id="interaction-with-cfg"><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e902d5cf-0ffa-4b7b-b21c-191f19e2a52e/width=525/e902d5cf-0ffa-4b7b-b21c-191f19e2a52e.jpeg" />⚙️ Interaction with CFG</h3><ul><li>CFG 1–3 → fine lines and human detail.</li><li>CFG 5–6 → refined sketch/cartoon.</li><li>CFG 8–10 → saturated poster-like look.</li></ul>👉 In short: Velvette doesn’t collapse at high values—it transforms them into additional styles. This sets it apart from other checkpoints and allows free experimentation.<hr /><h3 id="quick-guide">⚙️ Quick Guide</h3><ul><li>Steps → 10–30 (optimal: 14–30)</li><li>CFG → 2.5 for smoothness | 3+ for stronger definition</li><li>Samplers → DPM++ 2M Karras, Euler a</li><li>Resolution → 1024×1024+ (1152×1344 and 1536×896 work very well)</li><li>Upscalers → R-ESRGAN 4x+ Anime6B / 4x-UltraSharp</li><li>VAE → integrated</li></ul><hr /><h3 id="useful-tags">🏷️ Useful Tags</h3>Basic styles: <code>pvdc_il, pvdc_stl, sketch style, pastel style, cartoon style, thick lines, hatched style, comic style, detailed style, fabulosity style, cutecore, horrorcore, glitchcore, cutelatte, anime, shiny</code>Examples:<ul><li><code>(sketch style + thick lines + hatched style)</code></li><li><code>(comic style + detailed style)</code></li><li><code>(pastel style + fabulosity style)</code></li></ul><hr /><h3 id="notes">🧾 Notes</h3><ul><li>Providence was reduced to avoid redundancies.</li><li>Styles like oil painting were removed (they already exist in other checkpoints).</li><li>Maintains light NSFW support (less intense than Utopia).</li></ul><hr /><h3 id="prompting-tips">🎨 Prompting Tips</h3>General positives: <code>(masterpiece), (best quality), (ultra-detailed), perfect composition, intricate details, official wallpaper</code>By theme:<ul><li>Anime/Cartoon → <code>anime, cutecore, shiny skin</code></li><li>Realistic → <code>(REALISTIC:1.15), detailed eyes, realistic lighting</code></li><li>Experimental → <code>glitchcore, watercolor, abstract, dark theme</code></li></ul>Recommended negatives: <code>bad quality, worst quality, watermark, text, bad hands, missing fingers, logo, simple shading, blurry, low contrast</code>Extras: <code>flat color:2.5, washed out:2.1, low saturation:2.5</code><hr />📌 In conclusion: Velvette-XL is a versatile and creative checkpoint, capable of alternating between cartoon, sketch, and detailed styles. Thanks to the use of negative triggers, it offers unique flexibility to control the intensity of its base style and unlock additional visual modes.<hr /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/ee1897cc-b50b-42de-91df-5d7e771280da/width=525/ee1897cc-b50b-42de-91df-5d7e771280da.jpeg" /><span style="background-color:rgb(26, 27, 30);color:rgb(222, 222, 222);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">Thanks so much for your support! <span style="background-color:rgb(26, 27, 30);color:rgb(250, 82, 82);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">♥

Files with SHA256: f6d799b9447e2100a84a5ef33d37e99a1e0a79adfa3087c63b98f8f7f16feb55