Coffmeteor's picture
Mirror published Civitai release: Storyflow Three-Frame In-Context
6ee9545 verified
Raw History Blame Contribute Delete
2.46 kB
{
"model_id": 2891699,
"version_id": 3269328,
"title": "Storyflow Three-Frame In-Context",
"source_model_name": "Krea2 StoryFlow — 3-Frame In-Context LoRA",
"source_version_name": "v0.1",
"base_model": "Krea 2",
"type": "LORA",
"source_url": "https://civitai.red/models/2891699?modelVersionId=3269328",
"trained_words": [],
"verified_usage": [],
"source_model_description_html": "<h2>原理</h2><p>参考FLUX In-Context LoRA训练方案训练的Krea2 三帧故事板,仍属于实验中 。</p><p>把三个连续状态放在同一张纵向画布中联合生成,让模型利用画布内上下文保持人物、服装、场景和道具连续,同时表现动作、位置或镜头变化。</p><h2>使用方法</h2><h3>推荐参数:</h3><pre><code>Base: Krea2 Turbo BF16\n\nLoRA strength: 0.85(可尝试 0.7–1.0)\n\nResolution: 512×1536\n\nSteps: 8\n\nCFG: 1.0\n\nSampler: Euler Ancestral\n\nScheduler: Beta</code></pre><h3>提示词格式:</h3><pre><code>[SHOT-SEQUENCE] 三帧共同保持的人物、服装、场景、道具和光线。\n\n[FRAME-1] 初始状态。\n\n[FRAME-2] 中间状态。\n\n[FRAME-3] 最终状态。</code></pre><p>建议先在 <code>512×1536</code> 生成(训练分辨率),再把三格分别超分。不建议直接生成超高纵向分辨率。</p><h3>当前问题</h3><ul><li><p>小幅凝视、表情和手势变化有时不明显。</p></li><li><p>复杂手部与道具交互可能出错或重复道具。</p></li><li><p>强度过高时动作会变保守,后两格可能过于相似。</p></li><li><p>远景人物的脸部细节有限。</p></li></ul><p></p>",
"source_version_description_html": null,
"permissions": {
"allowNoCredit": true,
"allowCommercialUse": [
"Image",
"RentCivit",
"Rent",
"Sell",
"SellMerge"
],
"allowDerivatives": true,
"allowDifferentLicense": true
},
"files": [
{
"id": 3153106,
"name": "krea2-meridian-storyboard-v003-final.safetensors",
"type": "Model",
"path": "storyflow/three-frame-in-context/krea2-meridian-storyboard-v003-final.safetensors",
"sha256": "0c0a6666ea4f5f95c1be13419e146f3a07addac9f1f74d6c3b29f2772f6f744f",
"size_kb": 111673.046875,
"metadata": {
"format": "SafeTensor",
"fp": "bf16",
"isRequired": false
},
"download_url": "https://civitai.red/api/download/models/3269328?fileId=3153106"
}
]
}