EasyGenerate

Generate AI images with one python file!

Rainy neon Tokyo

Island in a bottle

Cathedral ballerina

Product poster

_{All four generated on a laptop CPU with the default settings — see Image_Tests/.}

Prompts used for these samples (click to expand)

1 · Rainy neon Tokyo

A futuristic Tokyo street at night during heavy rain, neon signs reflected in puddles, pedestrians with transparent umbrellas, flying drones overhead, dense atmospheric fog, cyberpunk aesthetic, extremely detailed, wide-angle composition.

2 · Island in a bottle

A glass bottle containing a miniature tropical island with palm trees, a waterfall, tiny birds, and glowing fireflies, sitting on a wooden desk beside an open notebook, magical realism, macro photography, soft window light.

3 · Cathedral ballerina

A ballerina performing mid-air inside an abandoned cathedral with shattered stained-glass windows, dramatic volumetric light rays, flowing fabric, realistic anatomy and hands, dynamic motion blur, high-detail cinematic composition.

4 · Product poster

A clean minimalist product poster for a fictional smartwatch called Aster One, centered on a white background, premium Apple-style advertising design, with the headline Aster One and the tagline Time, refined rendered as perfectly readable typography.

All four used the built-in default negative prompt. Note that SD 1.5 cannot reliably render text — the poster prompt asked for readable lettering and got shapes instead.

Single-file, CPU-only image tools. Run one, type a prompt, get a PNG. No GPU, no web UI, no API keys — just Python and patience.

There are two scripts, and they are not the same thing under the hood:

Script	Job	Model
generateimage.py	Text → image. Describe something, get a new picture.	Lykon/dreamshaper-8 (an SD 1.5 fine-tune)
EDIT-PHOTO/editimage.py	Image + instruction → image. Point it at a picture and tell it what to change.	timbrooks/instruct-pix2pix

Both use Hugging Face Diffusers, and both are separate downloads — using the editor does not reuse the generator's weights.

---

Requirements

• Python 3.9+
• ~8 GB RAM free (models are loaded in float32)
• ~5 GB disk per model, so ~10 GB if you use both (downloaded once, cached)

Install

git clone https://github.com/lllons/EasyGenerate.git
cd EasyGenerate

python -m venv .venv
# Windows
.venv\Scripts\activate
# macOS / Linux
source .venv/bin/activate

pip install torch diffusers transformers accelerate safetensors pillow

On Linux, plain pip install torch pulls the ~2.5 GB CUDA build. For a much smaller CPU-only wheel: pip install torch --index-url https://download.pytorch.org/whl/cpu

One install covers both scripts — the editor needs no extra packages.

Generating images

python generateimage.py

The first run downloads the model (a few GB) into your Hugging Face cache — subsequent runs start straight away.

You'll be asked for two things:

1. Prompt — what you want to see.
2. Negative prompt — what you don't want. Press Enter to accept the default (blurry, low quality, distorted, extra limbs, bad anatomy, text, watermark).

Output is saved next to the script as phto1.png, phto2.png, and so on — existing files are never overwritten.

Example

Enter your PROMPT:
> a lighthouse on a rocky coast at dusk, dramatic clouds, cinematic lighting

Enter your NEGATIVE PROMPT (leave blank for default):
>

Expect roughly 1–3 minutes per image on a modern laptop CPU at the default settings.

Tinkering

All the knobs live in section 3 of generateimage.py. Edit the constants and re-run.

Setting	Default	What it does
WIDTH / HEIGHT	512	Output size. Must be multiples of 8. SD 1.5 is trained at 512×512 — going much higher tends to produce duplicated limbs and heads, and costs a lot more CPU time.
STEPS	12	Denoising steps. More steps = cleaner detail, linearly slower. 12 is fast-and-rough; 20–30 is the usual sweet spot.
GUIDANCE_SCALE	7.5	How strictly the model obeys the prompt. Low (3–5) is loose and creative, high (10–15) is literal but can look overcooked.
SEED	16	Locks randomness so the same prompt gives the same image. Set to -1 for a different result every run.
OUTPUT_PREFIX	"phto"	Filename prefix for saved images.

Reproducibility tip: keep SEED fixed while you iterate on wording, so you can tell whether a change came from your prompt or just from a new roll of the dice. Switch to -1 once you're happy and want variations.

---

Editing images — EDIT-PHOTO/

Same one-file, no-GPU deal, but starting from a picture you already have. This one runs InstructPix2Pix (timbrooks/instruct-pix2pix), a different model to the generator, so the first edit triggers its own multi-gigabyte download.

The important difference is what you type. InstructPix2Pix takes an instruction, not a description:

✅	Make it snow · Turn the sky purple · Give him a leather jacket · Make it look like a watercolour painting
❌	a snowy street · a purple sky at dusk, cinematic lighting

Describing the scene the way you would for generateimage.py will confuse it. Tell it what to do.

Run

cd EDIT-PHOTO
python editimage.py

Before running, put the image you want to edit in the EDIT-PHOTO/ folder and name it input.png — the script doesn't ask for a path. (To use a different file, change the INPUT_IMAGE constant in section 3.)

You'll be asked for two things:

1. Edit instruction — the change you want made.
2. Negative prompt — same default as the generator; press Enter to accept it.

Results are saved next to the script as edit1.png, edit2.png, and so on. As with the generator, existing files are never overwritten.

Enter your EDIT INSTRUCTION:
> make it winter, cover everything in fresh snow

Enter your NEGATIVE PROMPT (leave blank for default):
>

Tinkering

Section 3 of editimage.py, same as the generator — but with an extra dial and a couple of different defaults.

Setting	Default	What it does
INPUT_IMAGE	"input.png"	The source file, relative to the script. Change this instead of renaming your photos every time.
WIDTH / HEIGHT	512	The input is resized to this before editing, and the output comes back at the same size. Multiples of 8 only. InstructPix2Pix has no separate size argument.
STEPS	5	Deliberately low so you get something back quickly on CPU. It looks rough. Raise it to 15–30 once you've found an instruction that works.
GUIDANCE_SCALE	7.5	How hard it pushes toward your instruction. Raise it if the edit isn't happening at all.
IMAGE_GUIDANCE_SCALE	1.5	How hard it pulls back toward the original. Range roughly 1.0–2.5. This is the dial that matters.
SEED	-1	Random every run by default. Set an integer to reproduce an exact result.
OUTPUT_PREFIX	"edit"	Filename prefix for saved images.

Tips

• The two guidance scales fight each other, and tuning is mostly about balancing them. Edit too weak or ignored? Raise GUIDANCE_SCALE or lower IMAGE_GUIDANCE_SCALE. Whole image mangled and unrecognisable? Do the opposite. Move one at a time.
• Lock the seed before you tune. The default -1 re-rolls every run, so you can't tell whether a change came from your settings or from luck. Set SEED to any integer while experimenting and switch back to -1 for variations.
• Bump STEPS for anything you'll keep. 5 steps is a preview, not a result.
• One change at a time. Compound instructions ("make it snow and add a red car and change it to night") tend to half-do all three. Run the output back through as the next input.png instead.
• Non-square inputs get squashed to WIDTH × HEIGHT. Crop to your intended aspect ratio first, or set the constants to match (both multiples of 8).
• It pairs with the generator. Copy a phto*.png you like into EDIT-PHOTO/ as input.png and iterate on it from there.

---

Running on a GPU

Both scripts are pinned to CPU. If you have a CUDA card, change these two lines in whichever one you're running:

torch_dtype=torch.float16   # was torch.float32
...
pipe = pipe.to("cuda")      # was "cpu"

...and update the generator device to match:

generator = torch.Generator("cuda").manual_seed(SEED)

On Apple Silicon, use "mps" instead of "cuda" and keep float32.

---

Notes

• The NSFW safety checker is disabled (safety_checker=None) in both scripts. That's a deliberate choice for local experimentation — worth knowing before you point anyone else at this, and worth a second thought before feeding in photos of real people.
• Model weights are cached in ~/.cache/huggingface/hub (%USERPROFILE%\.cache\huggingface\hub on Windows). Both models live there; delete the folder to reclaim the space.
• If the process is killed partway through, you're almost certainly out of RAM. Close browser tabs, or drop the resolution.
• Add *.png to your .gitignore unless you actually want your generations in version control — note that this would also ignore EDIT-PHOTO/input.png.

License

This project is released under the MIT License.

Model weights are governed separately: dreamshaper-8 by the CreativeML Open RAIL-M license, and instruct-pix2pix by its own model card terms.