下载与 SD1.5 匹配的官方 LCM-LoRA，把 MODEL 和 CLIP 双线接入 LoraLoader，用同一提示、种子、512×512、8步及sgm_uniform采样设置实际生成强度0／1两张图。LCM 是加速适配器，只有 UNet 权重，本例不能证明 CLIP 强度改变文本编码。
本次操作
- 确认家族和权重：使用SD1.5 FP16与 lcm-lora-sdv1-5，记录两者来源、revision、SHA和许可证；避免SDXL混入。
- 接好双线：基模 MODEL、CLIP分别连LoraLoader同名输入；输出MODEL给KSampler，CLIP给两个TextEncode。
- 固定变量对照：初次256×256、4步normal调度产生色带，原稿保留。按官方例子修正为512×512、8步、cfg=1、lcm/sgm_uniform，并加ModelSamplingDiscrete(lcm)。固定种子20261024，只对照LoRA强度0和1。
- 区分LoRA类型：本例LCM仅含UNet权重，CLIP线应接正确，但CLIP强度没有已证实的视觉变化；不要推广成所有风格LoRA结论。
结果与限制
- 两次原生ComfyUI推理均成功，并保留真实不同输出。
- 旧端口接反和缺LoRA负例继续保留。
- 初次采样配置有明显色带；修正调度后的真实512×512结果已检查，可见蓝色陶瓷杯。
- 低分辨率链路验证不等于 LoRA 风格美学评测；使用LCM时按其官方说明选择采样设置。
base-0
output-0.png · dfd5cd5954633bfdvalidation.json
{
"success": true,
"seconds": 12.035484125000949,
"outputSha256": "dfd5cd5954633bfd792cf24c31db7bbe148ac28c9f562df738cdb0161ae5d7af",
"strengthModel": 0,
"strengthClip": 0,
"actualDiffusion": true,
"loraType": "LCM acceleration adapter, UNet weights only; CLIP strength does not demonstrate a text-encoder effect",
"size": [
256,
256
],
"steps": 4,
"seed": 20261024
}lcm-lora-1
output-0.png · 9a8c3d1389beff7cvalidation.json
{
"success": true,
"seconds": 9.040923790998932,
"outputSha256": "9a8c3d1389beff7c93fcdf1ddac4ca0b71d74fbe76a4ce1d370981852ab65669",
"strengthModel": 1,
"strengthClip": 1,
"actualDiffusion": true,
"loraType": "LCM acceleration adapter, UNet weights only; CLIP strength does not demonstrate a text-encoder effect",
"size": [
256,
256
],
"steps": 4,
"seed": 20261024
}correct-lcm-0
output-0.png · c7bf1f5c2daaee3fvalidation.json
{
"success": true,
"seconds": 58.2011690420004,
"outputSha256": "c7bf1f5c2daaee3f820495255725d328efedab5b89fc319f9a442964996610b4",
"strengthModel": 0,
"strengthClip": 0,
"actualDiffusion": true,
"loraType": "LCM acceleration adapter, UNet weights only; CLIP strength does not demonstrate a text-encoder effect",
"size": [
512,
512
],
"steps": 8,
"sampler": "lcm",
"scheduler": "sgm_uniform",
"seed": 20261024
}correct-lcm-1
output-0.png · b9b7d84642f8c03bvalidation.json
{
"success": true,
"seconds": 45.693677874998684,
"outputSha256": "b9b7d84642f8c03bb4cf6fc1d3363915c70d9bdeaebfe1c1f73efcfdb8896236",
"strengthModel": 1,
"strengthClip": 1,
"actualDiffusion": true,
"loraType": "LCM acceleration adapter, UNet weights only; CLIP strength does not demonstrate a text-encoder effect",
"size": [
512,
512
],
"steps": 8,
"sampler": "lcm",
"scheduler": "sgm_uniform",
"seed": 20261024
}