This is a Latent Consistency model (Kind of*). Read Instructions below.
This model uses the LCM sampler. It is available in AUTOMATIC1111/stable-diffusion-webui through the [animatediff extension] or the [sd-webui-lcm extension].
In case LCM is new to anyone. Simply put. It allows for generating images in just 4-8 steps.
My model though works best at 1.5-2 CFG and 6-12 steps.
Super fast generation.
generates 768x1024 without any issues.
For the higher resolutions I have used the [Kohya Hires.fix extesnion] using the parameters uploaded to images below.
Recommended but not necessary negative::
(worst quality, low quality, low res:1.0), (bad anatomy, deformed, long body, tall, extra leg, extra arm, extra limb, duplicate, copy, clone, twin:1.0), (3d, render, flat art, drawing, illustration, sketch, painting, fanart, anime, comic, fine-art, video game, sculpture, blender \(software\):1.0), (monochrome, desaturated. sepia, vignette, dull colors:1.0)
Recommended suffix for positive
(best quality, hi res, absurd res:1.0)For realism (model is versatile and uses custom text encoder aka. CLIP)
(real, photorealism, real life, photography \(artwork\):1.0)- The model is a hybrid model that uses some of the LCM weights without them being fully trained. So using higher steps works fine and can be beneficial if you want to experiment using CFG rescale extension or CADS extension
Description
FAQ
Comments (5)
Hi, can you perhaps add some more showcase examples what your model can create?
You added pictures of one same looking character.
So this may lead into people thinking a LoRA has been added and the entire model creates this character and is more a LoRA than a model :)
Thx in Advance
The current showcase is primarily to provide prompt parameters.
It can create what you want it to create basically.
All boils down to prompt. As you might have noticed there is no recommendation for embeddings or any such thing.
The model however does perform better when using prompting that have words that are same as e621.net tags.
I chose this route because it makes the model capable of being free from the need to use embeddings or LoRA with the way I merged it.
I will upload more when I have time. This model was just impulse upload when I managed to create a workaround to hybridize a normal model into a LCM model without fine tuning.
Hello SILV3RC01N, thx for this checkpoint, can you change the Base Model SD 1.5 to Base Model SD 1.5 LCM. Please, this is to find you more easily. ^.^)
Has been Done
Muito bem otimizado, impressionante