<h2 id="yuan">逆光通透二次元画风</h2><p></p><p>建议使用MOSS-Image-heal</p><p>细线条、干净的皮肤上色、头发上的细碎高光和脸上的小块光斑，适合逆光、波光粼粼的海面、华丽服装和双人贴脸这类画面。</p><h3>推荐设置</h3><ul><li><p>触发词：<code>xiaoyue</code>（中文提示词用 <code>xiaoyue画风</code>），放在提示词最前面</p></li><li><p>LoRA 强度：1.0（测试都用的 1.0）</p></li><li><p>底模：MOSS-Image-heal</p></li><li><p>采样：euler / simple，50 步，CFG 3，shift 5</p></li><li><p>建议加跳层引导 SkipLayerGuidanceDiTSimple：single 22，0～50%</p></li><li><p>尺寸：横图 1536×1024 测试最多</p></li><li><p>负面：worst quality, bad anatomy, watermark, jpeg artifact, blurry</p></li></ul><h3>示例提示词</h3><p>xiaoyue. Two girls leaning their heads together on a sunny balcony. The girl on the left has short black hair and amber eyes and makes a peace sign with a wide smile. The girl on the right has long silver hair with a gold hairpin and violet eyes and looks at the viewer with a gentle smile, wearing an ornate white and gold dress. Bright sky with large white clouds, soft light flares.</p><p>英文标签、英文自然句、中文标签、中文自然句都参与了训练，四种写法都可以用。</p><h3>训练信息</h3><ul><li><p>底模：MOSS-Image-heal，用 KohakuMoss 训练</p></li><li><p>数据：58 张，其中 37 张为原图，21 张是用 v1 生成后人工挑选的补图</p></li><li><p>标注：PixAI 标签 + GBC 自然句，人工逐张核对，四种 caption 混训</p></li><li><p>rank 32 / alpha 128，学习率 1e-4 恒定，RTX 5090 每步 2 张</p></li><li><p>本版本为第 3400 步，每张图约学习 117 遍</p></li></ul><h3>已知限制</h3><ul><li><p>SFT 自带的光影比较强，暗场景里仍会带一些 SFT 的电影感暖光</p></li><li><p>大面积高亮的场景容易偏白、发淡，提示词里少写过曝，多写深色的物件和阴影会好一些</p></li><li><p>眼睛的虹膜细节比原画简单，脸也偏圆一些</p></li><li><p>这一档放在 heal 上用时颜色会偏灰偏淡，建议在 SFT 上使用</p></li></ul><hr /><h2 id="sunlit-anime-glow-style">Sunlit Anime Glow Style</h2><p>Thin linework, clean skin shading, fine highlights in the hair and small patches of light on the face. Works best with backlighting, glittering water, ornate costumes and close two-character shots.</p><h3 id="recommended-settings">Recommended settings</h3><ul><li><p>Trigger: <code>xiaoyue</code> (<code>xiaoyue画风</code> for Chinese prompts), at the start of the prompt</p></li><li><p>LoRA strength: 1.0 (all tests used 1.0)</p></li><li><p>Base: MOSS-Image-heal</p></li><li><p>Sampler: euler / simple, 50 steps, CFG 3, shift 5</p></li><li><p>Skip layer guidance recommended (SkipLayerGuidanceDiTSimple): single 22, 0–50%</p></li><li><p>Size: mostly tested at 1536×1024 landscape</p></li><li><p>Negative: worst quality, bad anatomy, watermark, jpeg artifact, blurry</p></li></ul><h3 id="training">Training</h3><ul><li><p>Trained on MOSS-Image-heal with KohakuMoss</p></li><li><p>58 images: 37 original images and 21 images generated with v1 and hand-picked</p></li><li><p>Captions: PixAI tags and GBC natural language, checked by hand, in English and Chinese</p></li><li><p>Rank 32 / alpha 128, constant LR 1e-4, 2 images per step on an RTX 5090</p></li><li><p>This version is step 3400, about 117 passes per image</p></li></ul><h3 id="known-limits">Known limits</h3><ul><li><p>SFT's own lighting stays fairly strong, and dark scenes keep some of its warm cinematic look</p></li><li><p>Very high-key scenes can turn pale. Fewer overexposure terms and more dark objects and shadows help</p></li><li><p>Eyes have simpler irises than the source art, and faces lean rounder</p></li><li><p>On heal this checkpoint turns greyer and paler, so SFT is recommended</p></li></ul><p></p>

Files with SHA256: 9dd2e8e7e535bf5e6d4521f2a5f035beeaff4f97797dce10e3a8a2b6663a1e50