---
language:
  - en
license: other
tags:
  - nanites
  - retro-web
  - text-generation
  - text-to-image
base_model:
  - unsloth/gemma-4-E4B-it
  - stable-diffusion-v1-5/stable-diffusion-v1-5
library_name: peft
pipeline_tag: text-generation
---

# nanites retro web designer (bundle of two LoRAs)

![A pre-Y2K web designer's desk](media/retro_web_designer.jpg)

One repo, two fine-tuned LoRA adapters that turn a plain request into
a complete pre-Y2K (1994-1999) web page **with materialized image assets**:

- `web/adapter/` — a **Gemma-4-E4B-it** QLoRA (bilingual 4B-class) fine-tuned
  on real archived pages (GeoCities / Tripod / Angelfire / personal sites)
  to emit era-correct HTML (`<center>`, `<font>`, `<marquee>`, table frames...).
- `image/adapter/` — a **Stable Diffusion 1.5** LoRA fine-tuned on GIFCities
  era GIFs (buttons, banners, dividers, backgrounds, clipart) so every
  `src`/`background` in the generated page resolves to an on-era asset.

The two are loaded together in one process (fine on 8 GB) to produce a
fully local, self-contained page.

## Trigger tokens

| expert | token |
|---|---|
| web coder | `retro web 1999` |
| image asset | `retroweb90s` |

## Quick start (web coder)

```python
import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "unsloth/gemma-4-E4B-it"           # the base the LoRA trains against
model = AutoModelForCausalLM.from_pretrained(
    base, torch_dtype=torch.bfloat16, device_map="cuda").to("cuda")
model = PeftModel.from_pretrained(model, "./web/adapter")
tok = AutoTokenizer.from_pretrained(base)
prompt = "retro web 1999, make me a website about my cat.\n<html>"
ids = tok(prompt, add_special_tokens=True)["input_ids"]
out = model.generate(
    torch.tensor([ids]).to("cuda"), max_new_tokens=700,
    do_sample=True, temperature=0.8, top_p=0.9)
print(tok.decode(out[0][len(ids):], skip_special_tokens=True))
```

The adapter config ships with local base paths blanked; pass your own
base explicitly. For the full generate flow (page + synthesized assets)
run the repo's `generate.py --assets synth` with these adapters mounted
at `run/adapter/<expert>`.

## Quick start (image asset)

```python
import torch
from peft import PeftModel
from diffusers import StableDiffusionPipeline

base = "stable-diffusion-v1-5/stable-diffusion-v1-5"
pipe = StableDiffusionPipeline.from_pretrained(
    base, torch_dtype=torch.float16, safety_checker=None).to("cuda")
pipe.unet = PeftModel.from_pretrained(pipe.unet, "./image/adapter")
img = pipe("retroweb90s, a small 1990s web button graphic named boost.gif, "
           "theme visiting, vintage 90s website clipart").images[0]
```

## Samples

`samples/` holds pages generated end-to-end by this bundle (LLM page + synth assets): goat soap, honda civic, personal notes, cat homepage.
Each `*.html` is openable directly; every image ref points at the sibling
file next to it.

## Training

- **web**: QLoRA r16 on ~thousands of real 1994-99 archived pages
  (`experts/retro_web_designer/data/`); NF4 quantized linears + CPU embeddings on RTX 4070 8 GB, raw continuation (no distillation).
- **image**: LoRA r16 on ~1500 curated GIFCities era GIFs (filtered, despeckled, deduped) at 256px (`experts/retro_image_designer/data/`); SD1.5 UNet, fp16.

## License

- The web adapter is a derivative of **Gemma 4** weights (Gemma Terms of
  Use apply to the base and derivatives).
- The image adapter is a derivative of **Stable Diffusion 1.5** UNet
  weights (CreativeML Open RAIL-M applies).
- Training pages are historical archives reproduced for research/art;
  training assets remain the property of their original creators.