Nostalgic Dream 1.0

Please share the images you generate so I can make future improvements with more precision.

Leave your like and help me popularize this model.

The main focus of this model is to try to work well with all body parts, including feet and hands, even eyes, and especially scenarios without the need for complex adjustments and descriptions (except for hands and feet, in some cases you will need to be precise when describing the positioning and pose).

(recommendation) USE SDLX VAE: DOWNLOAD HERE (prefer to use in fp32)

Recommendations:

Scheduler: DPM++ 2S

Steps: between 30 and 35

CFG scale: 4 or 5

VAE precision: fp32

Image size: try something from 800x800 (I think the engine I use influences it a bit, anyway, do some tests here, the image may be distorted if you make it too small).

Sometimes there may be an angelic halo due to the training images. Include this in the negative prompt to mitigate this.

Suggested negative prompt:

score_6, score_5, score_4, worst quality, low quality, text, censored, deformed, bad hands, malformed hands, blurry, (watermark), extra hands, deformed fingers, extra fingers, deformed legs, missing limbs, muscular, limb fusion, blending, angelic halo, halo

Checkpoint created with approximately 23,000 images selected based on quality, trying to maintain an average of 1,000 to 2,000 images with similar characteristics in order to increase focus on the poses, especially when it comes to hands and feet (it is very likely that a considerable part of this training content was also generated by AI, but I still tried to filter it as best as possible during two weeks of intense personal analysis). Even with this strategy, it can still cause headaches; The unwanted halos that may appear are due to the fact that most of the images used for training are from Azur Lane and Blue Archive, but I believe that they add a lot of quality to the model, and I don't feel like checking that right now. In any case, the model is a bit too bright, which I don't like very much, so I'll have some updates on that soon.