Velvette-XL ✨ — 「illustrious-SDXL」

Velvette-XL ✨Velvette-XL is a refined version of a discarded project (an unreleased checkpoint), recovered and reconstructed with the intention of exploring hybrid and versatile styles for my personal taste.It is based on a fusion of 2D styles, ranging from stylized sketches to semi-realistic detailed illustrations. Its main focus is line versatility: it can produce both illustrations with defined strokes and lineless art, always with a high level of detail.Velvette was born from the previous concept — Providence —, an experimental LoRA that combined multiple styles through tags. The same philosophy applies here: a lightweight, flexible, and adaptable model, capable of adjusting to the desired style without losing consistency, while offering a more complete creative experience.V4 | Noblesse FOCUSI’ve been using sd-mecha since v3, and for v4 I ended up working with far too many variables. I was used to SuperMerge, so I had to completely rethink my merging workflow—switching 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 resets (the very first build is also merged in here), and there are more than twenty internal iterations—some better, some worse. At first I treated it as just an experiment, but once I found a solid workflow, I pushed it as far as possible and refined it to the limit.This version includes a large number of merges. There’s no recipe because even I don’t remember how many components went into it—LoRAs, models, images… everything. It blends furry, hentai, 3D, 2D, anime, and more. In short, it’s a complete model designed to preserve the maximum amount of information available.V3 | ALPHAALPHA is actually an incomplete version (it’s a fusion of v1 & v2), the idea was to merge it with IL v2.0 or v3.0 (which should be out soon), also rouwei v7 although I’ve been playing with this for days and ended up gaining affection for it so I haven’t finished it, alpha has a sister version that includes pony diffusion, it’s just that I haven’t tested it much and I’ll probably add other fusions. (I’m a bit disappointed with Civitai, but I recently managed to merge models into free colab accounts and that lightens my wallet)V2 | EPSThere 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 cmd2 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)/🎛️ Behavior with CFG and StepsCFG 1–4 → Faithful to applied LoRA, cleaner lines.CFG 4–7 → Velvette’s base style takes over, sketch-like look.CFG 7–8+ → More saturated and cartoonish appearance, similar to using a style LoRA at strength 1.5–2.📌 Practical tip:If you’re not using LoRAs at low CFG, add:Positive → hdr:2.5Negative → flat color:2.5This adds extra detail even at CFG 1–2, reaching quality comparable to Utopia v4 at CFG 4.🔹 Step PerformanceMinimum useful → 6 steps at CFG 1.Optimal → 12–30 steps.⚙️ Quick GuideSteps → 10–30 (optimal: 12–30)CFG → 1.5 for smoothness | 3+ for stronger definitionSamplers → DPM++ 2M Karras, Euler a, turbo, etc.Resolution → 1024×1024+ (1152×1344 and 1536×896 work very well)Upscalers → R-ESRGAN 4x+ Anime6B / 4x-UltraSharp🏷️ Useful TagsBasic styles: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, shinyExamples:(sketch style + thick lines + hatched style)(comic style + detailed style)(pastel style + fabulosity style)🎨 Prompting TipsGeneral positives:(masterpiece), (best quality), (ultra-detailed), perfect composition, intricate details, official wallpaperBy theme:Anime/Cartoon → anime, cutecore, shiny skinRealistic → (REALISTIC:1.15), detailed eyes, realistic lightingExperimental → glitchcore, watercolor, abstract, dark themeRecommended negatives:bad quality, worst quality, watermark, text, bad hands, missing fingers, logo, simple shading, blurry, low contrastExtras:flat color:2.5, washed out:2.1, low saturation:2.5V1.1 | DeletedWell, 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).V1 | NormalThe 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.)🎛️ Behavior with CFG and StepsCFG 1–3 → Faithful to applied LoRA, cleaner lines.CFG 4–6 → Velvette’s base style takes over, sketch-like look.CFG 7–8+ → More saturated and cartoonish appearance, similar to using a style LoRA at strength 1.5–2.📌 Practical tip:If you’re not using LoRAs at low CFG, add:Positive → hdr:2.5Negative → flat color:2.5This adds extra detail even at CFG 1–2, reaching quality comparable to Utopia v4 at CFG 4.🔹 Step PerformanceMinimum useful → 8 steps at CFG 1.Optimal → 14–30 steps.High CFG → experimental, depending on tags and style mixing.💡 Hidden Control with Cutelatte NegativesA 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⚙️ Interaction with CFGCFG 1–3 → fine lines and human detail.CFG 5–6 → refined sketch/cartoon.CFG 8–10 → saturated poster-like look.👉 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.⚙️ Quick GuideSteps → 10–30 (optimal: 14–30)CFG → 2.5 for smoothness | 3+ for stronger definitionSamplers → DPM++ 2M Karras, Euler aResolution → 1024×1024+ (1152×1344 and 1536×896 work very well)Upscalers → R-ESRGAN 4x+ Anime6B / 4x-UltraSharpVAE → integrated🏷️ Useful TagsBasic styles: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, shinyExamples:(sketch style + thick lines + hatched style)(comic style + detailed style)(pastel style + fabulosity style)🧾 NotesProvidence was reduced to avoid redundancies.Styles like oil painting were removed (they already exist in other checkpoints).Maintains light NSFW support (less intense than Utopia).🎨 Prompting TipsGeneral positives:(masterpiece), (best quality), (ultra-detailed), perfect composition, intricate details, official wallpaperBy theme:Anime/Cartoon → anime, cutecore, shiny skinRealistic → (REALISTIC:1.15), detailed eyes, realistic lightingExperimental → glitchcore, watercolor, abstract, dark themeRecommended negatives:bad quality, worst quality, watermark, text, bad hands, missing fingers, logo, simple shading, blurry, low contrastExtras:flat color:2.5, washed out:2.1, low saturation:2.5📌 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.