Base on LCM LORA
I made this Version for accelerating Checkpoint to generate.
Only 2 MB !!
Generation Time = 1s per SD15 0.5K img and 3s per SD15 4K quality img.
Suitable to make Animation : and 10 batchs or more once a time of generation not impossible anymore.
Remove the influence of CLIP:
now you can link the node wherever you want , no more just put them at the end of you LORAs and hesitate if you should link the CLIP or not which make your nodes more complex.
[Usage : 1. Just like a normal LORA. 2. use lcm sampler. 3. CFG==1-2. 4. Steps==2-8]
基于最新的ADD训练模式，LCM LORA的推出大大加快的出图速度。
我基于最新的两个官方LORA模型，制作了这个微型的LORA版本。
仅仅只有 2 兆字节 大小。
生图时间：1秒钟一张512分辨率的SD15图像，3秒钟一张SD15 2K 高清大图。
快来做动画！：一批次100张也不过10s/步（显存18G就行），而LCM的特性让生成步骤只有2-8步，也就是20s-80s就能出一百张图！
此版本特别优化了CLIP层对于图像的影响：
现在，使用我的LORA，你可以随意的将这个24MB的LORA放在你的LORAs栈中的任意位置，而不必再因为需要避开CLIP层而感到迷茫，费时费力的将它放在最后，并且需要链接更复杂的节点了。随便放，放哪里都可以。
【使用方法：第一步：像随便哪一个LORA一样使用它. 第二步：采样方法选LCM（没看见的快更新你本地的平台版本。）第三步：调节CFG为1到2之间。第四步：采样步骤在2-8步之间。然后点击生成图片就可以了】
Description
2MB
FAQ
Comments (15)
Thank you for sharing. Please share generation data like cfg, number of steps etc.
Usage I wrote in green color
@Aseer I use black theme in my web browser 😀 some texts were barely visible, that confused me. Can we use it for normal SD models? thanks.
@godpunisher emm, can you use it successfully? CFG==1-2, Steps==6-8 Sampler==LCM and add my LORA. and yes, it is flexible to any normal SD15 checkpoint
@Aseer thank you 😂
Does this work with any checkpoint? What's different about this versus previously released LCM Loras?
You can use other samplers like Euler a as well, and push the steps up to 10. Also, I was getting horrible results with the original LCM LORA when I used highres fix but this seems to work fine, well done!
I noticed the horrible results occurs in certain checkpoints only. I am getting excellent results with HIRES 2.5x, DENOISE 0,4, with base image 512x648, LCM sampler, steps 6 or 7, with checkpoints like photonV1, runDiffusionFX and halcyon (realistic). epicRealism and epicPhotoGasm required 8 or more steps. For anime, kotosmix seems to suffer a lot. RevAnimated122 needs 8 steps or more.
[TIP] The speed is almost the same as the 128mb original LCM lora. In both, I noticed the horrible results occurs in certain checkpoints only. I am getting excellent results with HIRES 2.5x, DENOISE 0,4, with base image 512x648, LCM sampler, steps 6 or 7, with checkpoints like photonV1, runDiffusionFX and halcyon (realistic). epicRealism and epicPhotoGasm required 8 or more steps. For anime, kotosmix seems to suffer a lot. RevAnimated122 needs 8 steps or more. If your renderings get 'blurry' or 'strange', please increase STEPS.
效果非常好！
Just tried it here on Civitai, it works quite well.
Best settings to me were: cfg 1 and sde++2m karras with steps 12 ~ 13 ||| euler A 14 ~ 15
And speed is like just 1sec... heh
it's amazing for only 2mb! Kudos! :D
I've posted some results here with a review, but in case it vanishes:
I'm not a fan of the quality you get with ultra fast LCM generations, but this LoRA can also be used in more subtle ways to enhance generations with low but not super low sample step counts. 0.1 strength gives a subtle improvement with 8-10 steps DPM++ 2M Karras.
SDE version is on sdxl but can you make one for 1.5?
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.