<h1 id="starlit">Starlit</h1><p>One of the first SDXL models I've used is <a target="_blank" rel="ugc" href="https://civarchive.com/user/chillpixel">chillpixel</a> 's great <a target="_blank" rel="ugc" href="https://civarchive.com/models/143043/starlight-xl-animated">Starlight XL </a>and I continue to play with it at times. But since it looks like the author seems inactive, it doesn't look like it's going to be "ported" to newer models.</p><p>So I've created about 300 images with Starlight XL, selected a hundred of them and used them to train these loras in Flux, Chroma and Qwen. Think of this as an spiritual successor or homage to the original model (hence they name I've chosen).</p><h2 id="model-strength">Model Strength</h2><p>Strength typically works best between 0.5-0.8, although I find Flux more forgiving when using 1 as strength. In Chroma strength 1 tends to fail at eyes and hands in medium and general shots, perhaps because these loras are trained on sdxl synthetic images. I find the Qwen lightning lora overcooks the image when this lora is used at high strengths.</p><h1 id="samplers-and-prompting">Samplers and prompting</h1><p>As for parameters, Starlight XL "default" style is 2.5D anime (perhaps Korean style anime?) so I've trained it in illustration/digital art/cartoon/anime style of images. I can produce photos, with non realistic facial proportions (that's also kind of the point), but it's isn't tailored for it. So in Chroma it benefits from using the digital_media_(artwork) tag, as well as aesthetic_10 (I've tried 11 and it doesn't seem to overcook the image). In Qwen you use negative prompts, so you can enter photo, dslr to make sure it doesn't default to photographs. For Flux, it can be a matter of prompting with something like "A digital concept art about..." or something like that, as both Flux and Qwen tend to default to magazine covers.</p><p>For Qwen, I've always tested with res_multistep, beta and cfg 2.5.</p><p>For Flux, I've used euler or heun samplers, with either beta or sgm_uniform</p><p>For Chroma, I've tried several combinations. I've kept cfg between 3.2 and 4. Heun like samplers tend to do best in complex prompts, but I mostly stuck to the slower (but with fewer steps) res_2s. Res_multistep, gradient_estimation and euler samplers are also good (and faster) samplers. Beta, I think, produces good results.</p><h2 id="nsfw-images">NSFW images</h2><p>It was trained on topless and nudity, but the original lacks genitals. So it can easily do nudity and porn in Chroma (because the model can) and topless in Qwen and Flux, but it can't really produce full nudity, let alone sex, in neither Qwen nor Flux unless you mix it with other nsfw loras and since those are trained in photos, the results can be interesting and not in an aesthetically pleasing way. So if you're looking for NSFW, I'd stick to Chroma</p><p>Trigger words</p><p>They are pretty much unnecessary, and Qwen tends to write it if you write it, but they are:</p><p>Starl1t in Chroma</p><p>St4l1t in Flux</p><p>St4l1t as well in Qwen</p><p>St4rl1t in Krea 2</p><p>And without further Ado, I hope you enjoy it</p><h1 id="krea-2">Krea 2</h1><p>Since I had the dataset ready, I decided to train it for Krea 2. Needless to say, training Krea 2 on the output of an SDXL model isn't the best of ideas, but since I had it already, I decided I may as well publish it. For Krea 2, I recommend using it with a low strength, between 0.2-0.5 and definitely not for anything resembling photorealism. That way, the image still tends to generate the 2.5D head and eyes shapes without imprinting that much of sdxl skin textures.</p><p>For example, Krea 2 without loras</p><p><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/2c14b67c-786e-457f-aef6-8869eb112774/original=true/2c14b67c-786e-457f-aef6-8869eb112774.jpeg" /></p><p>With this lora, but at 0.5 strength</p><p><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/a40b5311-184a-4e33-96b8-8cdff584f85d/original=true/a40b5311-184a-4e33-96b8-8cdff584f85d.jpeg" /></p><p>But if we ramp it up to 0.9, we're pretty much generating an SDXL image with a better prompt adherence</p><p><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/7ead70c5-9cb0-49f6-9553-e1bfbe6f2f24/original=true/7ead70c5-9cb0-49f6-9553-e1bfbe6f2f24.jpeg" /></p>

Starlit

Files with SHA256: 0095f57e64280f78a7cc918ea8ef470da2e03917a6240f3f46bf1364eaa2b428