<h3 id="why-chaotique">Why Chaotique?</h3>Earlier Duchesse merges leaned toward a more defined house style, that signature look has been intentionally softened to allow for more versatility and creative freedom. A major focus during development was artist and style tag compatibility. I’ve spent a lot of time testing and mixing stylistic influences, and Chaotique is tuned to blend those tags smoothly. The lighting control is easily the best I’ve worked on so far and the model works well with every LoRA I’ve tested.Sample Image Settings<ul><li>Sampler: DPM++ 2M Karras / Euler A/ Restart</li><li>Steps: 25–35 (base), 40–50+ for Hi-Res Fix</li><li>CFG: 3–6</li><li>Clip Skip 2</li><li>For more information read <a target="_blank" rel="ugc" href="https://civarchive.com/articles/24298">The Chaotique Chronicles</a></li></ul>I included 20 sample images with no LoRAs <s>because I’m a narcissist obsessed with my own work</s> to highlight the model's versatility. It’s designed to perform well with long, complex prompts and short, minimal ones, consistently producing images that are vibrant, colorful, and visually engaging. And once again, special thanks to @Ziperto for all the advice/encouragement and for testing out some of the earlier versions of this model!Duchesse v3.0The latest and, in my opinion, best version of the model yet. Much better prompt adherence when using artist tags, a little more of the style from the first 2 versions with a lot more versatility. I would list all the checkpoints used in the merge but there's over a dozen and I totally lost track of them at some point amidst my usual chaos.I wrote an article showing how some tags change images and listing some of the artist tags I've been trying out with the model here: <a target="_blank" rel="ugc" href="https://civarchive.com/articles/22090">Duchesse Diaries</a>Edit: Huge oversight, I can't believe I forgot to mention this part. This model wouldn't have been possible for me to make without the help of @Ziperto. I learned so much from him about merging models that I was totally clueless about before. Check out his models <a target="_blank" rel="ugc" href="https://civarchive.com/models/1480704/mature-citron-il">Mature Citron IL</a> and <a target="_blank" rel="ugc" href="https://civarchive.com/models/1965943/amanesse-works">Amanesse Works</a>, they're both fantastic!I'll be putting an article together showing some of the artists tag and quality tag combinations I like with the model but I tried to include samples of my favorites so far.Most of my sample images were initally generated with Euler A 25 steps or DPM++ Karras 30 steps. then Hi-Res fix and detailers as usual. I went back to baking Liquid9745VAE by Liquidn2 into the model.Duchesse 2.0The latest version of Duchesse keeps the core style of the previous release while greatly expanding its versatility. I’ve highlighted in the sample images how some of the baked-in styles interact and blend.This version combines Duchesse v1.5, <a target="_blank" rel="ugc" href="https://civarchive.com/models/1162518?modelVersionId=1714002">Plant Milk Walnut</a>, <a target="_blank" rel="ugc" href="https://civarchive.com/models/1110783?modelVersionId=1670263">iLustMix v5.5</a>, and <a target="_blank" rel="ugc" href="https://civarchive.com/models/1353314?modelVersionId=1945082">Animij v4</a>. I went back to using <a target="_blank" rel="ugc" href="https://civarchive.com/models/1684460/illustrious-xl-v20?modelVersionId=1906483">Illustrious-XL-v2.0 VAE</a> for this version.Both DPM++ 2M Karras and Euler A work well. The sample images were generated with Euler A, 25 steps, followed by a HiRes fix and detailers for refinement.Duchesse 1.5 This update refines everything that made Duchesse 1.0 shine—without losing the signature style. The new build brings sharper details, improved body proportions, and noticeably better prompt adherence, all while keeping that elegant, painterly flair intact.A new VAE (Liquid9745VAE by Liquidn2) adds richer contrast and finer handling of small details, giving images more depth and presence. If you prefer the softer, lower-contrast look from Duchesse 1.0, you can easily swap back to the Illustrious-XL-v2.0 VAE.Think of this release as the same style, now with cleaner execution.Duchesse 1.0I love Illustrious but I was missing the feel of some Pony models so this is my attempt to bridge those worlds.Duchesse is a carefully crafted merge built on the Illustrious XL family of models with a stylistic infusion of LoRAs inspired by Pony models like Duchaiten v6. This checkpoint aims to capture the soft, elegant 2.5D anime style of Pony models while retaining the flexible structure and generation power of Illustrious models. The model attempts to lean away from the ultra-glossy plastic feel found in many Illustrious checkpoints, favoring instead a smoother, more painterly rendering with subtle detail and depth. Duchesse is ideal for users seeking a more naturalistic, artful finish with a touch of fantasy. Suggested Settings<ul><li>Sampler: DPM++ 2M Karras / Euler A</li><li>Steps: 25–35 (base), 40–50+ for detail</li><li>CFG: 4–7.5</li></ul>

Duchesse

Chaotique is a creative-first Duchesse merge focused on versatility, expressive prompting, and stylistic freedom. It handles both simple and complex prompts with ease, excels at artist and style tag blending, and offers exceptional lighting control. More anime-leaning than previous Duchesse models, its core style is intentionally restrained to adapt to a wide range of aesthetics and LoRAs. Built for experimentation, not rigidity.

duchesse_chaotique.safetensors

Chaotique

Mystique

v3.0

v2.0

v1.5

v1.0

Creator: ChumpleWhy Chaotique?Earlier Duchesse merges leaned toward a more defined house style, that signature look has been intentionally softened to allow for more versatility and creative freedom.&nbsp;A major focus during development was artist and style tag compatibility. I’ve spent a lot of time testing and mixing stylistic influences, and Chaotique is tuned to blend those tags smoothly. The lighting control is easily the best I’ve worked on so far and the model works well with every LoRA I’ve tested.I included 20 sample images with no LoRAs <s>because I’m a narcissist obsessed with my own work</s> to highlight the model's versatility.&nbsp;Sample Image Settings<ul><li>Sampler: DPM++ 2M Karras / Euler A/ Restart</li><li>Steps: 25–35 (base), 40–50+ for Hi-Res Fix</li><li>CFG: 3–6</li><li>Clip Skip 2</li></ul>

Files with SHA256: ce9b750c7be4234382db3e084feca1ec1aafebdd916c98f501bfddd5b45f850f