Stable Diffusion大模型

1. 首页
2. Ai应用下载
3. Stable Diffusion大模型

My goal was to push the boundaries even further than the last version. I've seriously toned down the surprise extra limbs! Now you can enjoy realistic, human-like figures that (fingers crossed) consistently stick to two arms and a reliable five fingers per hand. No more "Is that an elbow or modern art?" moments!

Trained using a different model apart from the base SD1.5 for better and more coherency between the images. V1 and V1-Alt were trained using the base SD1.5 model. Overall V1.5 creates better and more creative images than both those models but in some cases the image quality for V1 is more varied, mainly on the body stripes.

Note: Currently this model still doesn't have a tail. The ai has issues learning it. By V2 i will make sure the Ai is capable of creating the tail.

•Avatar Style V1.5 will introduce various amounts of fixes and improvements. Such as:

1. Better Teeth

2. Better Full Body

3. Better more accurate eyes

4. Better hair control

5. Better fingers and hands

6. Better skin coloration and less "color burn" or otherwise known as "Deep Fried"

Among other smaller improvements.

•As for issues with this new model there are a few still present but i hope to fix those by V2. Such as:

1. Missing Tail (In super rare cases i have seen the model do the tails but.... will improve that by V2)

2. Sometimes the eyes still wander to normal human eyes.

3. The fingers are sometimes crooked or deformed (might be fixable by negative prompts but I'll work on it)

4. The Navel is still human like and not Na'vi like.

5. Ears sometimes appear low quality.

6. lacking amount of stripes variety.

7. Minor teeth issues.

8. Body is still human like but at least it's better than previous models.

9. Minor lack of variety between head and face variations.

As for the question many are asking in their heads.... "Can it do NSFW?"

My answer is "Yes, to an extent. It has a bit of trouble with some NSFW activities but it can do naked bodies just fine." As for more NSFW activities I'm sure you can use a LORA, Hypernetwork or an Embedding.

Just add hyper-sd and vae

Sampling Method : LCM

Schedule type: Auto

CFG:1.5~2

Step : 8 (can up to 12~20 enhance the quality)