# Big Pot with NMFSAN

**No More Fucking Stupid Artist Name.**

Actually this model is broken, kinda.

Trained with custom embeddings with [naifu-diffusion](https://github.com/Mikubill/naifu-diffusion/blob/main/experiment/textual_inversion.yaml), aiming to solve the problem of poor style recognition because of CLIP.

[WandB report of this training](https://wandb.ai/crosstyan/BP_NMAN/runs/1o0l0dvs), which contains sample images and training log.

Actually I should not have trained this model based on BPModel since it has already
overfitted to the training set, but I did it anyway. As for the style
recognition, it is not as good as I expected, since most of styles I love are
already known without custom embeddings. Some tricky artists name now can be
recognized but Stable Diffusion DOES NOT HAVE MAGIC TO REPLICATE EVERYTHING.  I
need to review my training dataset but the poor result should not relate to CLIP
now.

Still mixing artists style is fun, which is better than merging various
models. Don't merge models.

## Usage

Use [bp_nman_e29.safetensors](bp_nman_e29.safetensors) with [custom embeddings](embeddings/agm.pt) in `embeddings` folder. Check [Textual Inversion](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Textual-Inversion) if
you don't know how to use embeddings.

## Cherry Picked Samples

Still playing with it. Still doing cherry picking.
More samples can be found in `images` folder. 

Simply replicating/imitating the style of artists is not the goal of this model.
I encourage you to try [Prompt Editing](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-editing) and [Alternating Words](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#alternating-words) when playing with the styles. Applying only one
style usually gives a bad result due to the overfitting.

![cat](images/00343-263196747_20230105142445.png)

```txt
by [alphonse::0.5] fuzichoco agm parted lips fox ears fox tails 1girl small breasts in serafuku pleated skirt red bow white thighhighs white hair two side up hair lime green eyes sitting in chair in New York city , cowboy shot
Negative prompt: (bad:0.98), (normal:0.98), (comic:0.81), (cropped:0.81), (error:0.81), (extra:0.81), (low:0.81), (lowres:1), (speech:0.81), (worst:0.81), blush, collage, nipples
Steps: 18, Sampler: DDIM, CFG scale: 6, Seed: 263196747, Size: 768x1280, Model hash: d7e48c33, Denoising strength: 0.7, Clip skip: 2, ENSD: 31337, First pass size: 0x0
```

![pink_girl](images/00152-1663577829_20230105135759.png)

```txt
[sketch:0.6] small breasts white hair 1girl pleated skirt pink eyes (presenting:1) sitting white shirt white thighhighs, as109 [fkey::0.5]
Negative prompt: (bad:0.98), (normal:0.98), (comic:0.81), (cropped:0.81), (error:0.81), (extra:0.81), (low:0.81), (lowres:1), (speech:0.81), (worst:0.81), blush, collage
Steps: 18, Sampler: DDIM, CFG scale: 9, Seed: 1663577829, Size: 768x1280, Model hash: d7e48c33, Denoising strength: 0.7, Clip skip: 2, ENSD: 31337, First pass size: 0x0
```

![宇宙人](images/01035-3125048569_20230105165845.png)
```txt
(painting:0.45) of solo loli fang 1girl with long hair in [astronaut spacesuit|sailor dress] dress pleated skirt headwear helmet standing, science fiction , upbody, bubble, satellite, by shion anmi, looking at viewer, apple and orange fruit
Negative prompt: (bad:0.98), (normal:0.98), (comic:0.81), (cropped:0.81), (error:0.81), (extra:0.81), (low:0.81), (lowres:1), (speech:0.81), (worst:0.81), collage
Steps: 18, Sampler: DDIM, CFG scale: 8, Seed: 3125048569, Size: 768x1280, Model hash: d7e48c33, Batch size: 4, Batch pos: 2, Clip skip: 2, ENSD: 31337
```

![walking](images/01295-2597684201_20230105174715.png)

```txt
[oil painting:0.75] [(sketch:1.25):0.65] profile side view expressionless cat ears loli 1girl blue eyes twintails blonde long hair in black serafuku bulletproof vest pleated skirt walking in [winter::0.75] [beach:city tree white blue flower:0.25] black glove, blood, fire, by (shion:1.1) fkey_alternative, cyborg, water splash, ice break
Negative prompt: (bad:0.98), (normal:0.98), (comic:0.81), (cropped:0.81), (error:0.81), (extra:0.81), (low:0.81), (lowres:1), (speech:0.81), (worst:0.81), collage
Steps: 18, Sampler: DDIM, CFG scale: 7, Seed: 2597684201, Size: 768x1600, Model hash: d7e48c33, Batch size: 4, Batch pos: 2, Denoising strength: 1, Clip skip: 2, ENSD: 31337, First pass size: 448x896
```