Various Trivial Diffuers Scripts for Personal Use

Collection of various local use convenience scripts for playing around with diffusers.

Level of Support Offered

I'll just offer a parable: "When selling a used car, the warrantly ends when the car moves under its own power and the merchant is able to see the tail lights."

Development Environment

I'm working on a Debian Bookworm box with an Nvidia RTX A4000.

If you have a functional Python 3.11 installation, uv run --locked should have you covered /wrt local setup.

Editorconfig

• UTF-8
• LF
• Managed whitespace
• Lines are 120 characters long

The line length is inherited by all other configs.

Formatting

• ruff

Linting

• ruff
• pymarkdown
• shellcheck

Refer to pyproject.toml for details on which linters are enabled and how they are set up.

Smoketest

The nature of these scripts does not allow for properly checking for correctness of operation and/or output. For this purpose a manual smoketest script is provided for assessing if any manual version bumps broke anything.

$ ./smoketest.sh

cogvideox.py

real 3m16.692s
user 1m46.756s
sys 1m22.098s

cogview3.py

real 2m47.431s
user 2m36.940s
sys 0m21.348s

flux-dev-16.py

real 6m10.513s
user 4m47.417s
sys 1m8.031s

flux-schnell-16.py

real 0m56.253s
user 0m26.351s
sys 0m13.899s

sa-10.py

real 1m6.092s
user 1m10.474s
sys 0m5.748s

sd-21.py

real 1m7.606s
user 1m1.965s
sys 0m15.734s

sd-30.py

real 1m27.871s
user 1m31.381s
sys 0m26.309s

sd-35.py

real 3m24.998s
user 2m26.810s
sys 0m32.012s

sd-35-turbo.py

real 1m37.374s
user 0m51.319s
sys 0m20.205s

sdxl-turbo.py

real 0m50.113s
user 1m3.175s
sys 0m14.755s

real 22m44.944s
user 17m42.589s
sys 5m0.139s

CI

• Python formatting: ruff
• Python linting: ruff
• MarkDown linting: pymarkdownlnt
• Reviewdog Shellcheck

Scripts Provided

Every script provided has default values for every parameter, so you can try them out without having to dive in deeper.

Stable Diffusion XL Turbo 16bit

• uv run --locked sdxl-turbo.py
• uv run --locked sdxl-turbo.py --help

A fast model, which produces something at a single step. Not terrible, not great.

Stable Diffusion 2.1 16bit

• uv run --locked sd-21.py
• uv run --locked sd-21.py --help

The memory use is not horrible and it's not too slow. The output is not too bad, but not great either.

Stable Diffusion 3.0 16bit

• uv run --locked sd-30.py
• uv run --locked sd-30.py --help

Medium quality and speed.

Flux.1 Schnell 16bit

• uv run --locked flux-schnell-16.py
• uv run --locked flux-schnell-16.py --help

Good quality and very fast. Somewhat memory hungry.

Flux.1 Dev 16bit

• uv run --locked flux-dev-16.py
• uv run --locked flux-dev-16.py --help

Good quality and somewhat slow. Memory hungry.

Stable Audio 1.0

• uv run --locked sa-10.py
• uv run --locked sa-10.py --help

Very mixed results. Not too slow. Not too memory hungry.

CogVideoX

• uv run --locked cogvideox.py
• uv run --locked cogvideox.py --help

Surprisingly good results. Surprisingly fast. Surprisingly not too memory hungry.

Stable Diffusion 3.5 16bit

• uv run --locked sd-35.py
• uv run --locked sd-35.py --help

Sorta fast and excellent results.

Stable Diffusion 3.5 Turbo 16bit

• uv run --locked sd-35-turbo.py
• uv run --locked sd-35-turbo.py --help

Hella fast for the quality and very good results.

Cogview3-plus 16bit

• uv run --locked sd-35-turbo.py
• uv run --locked sd-35-turbo.py --help

Not slow. Very saturated and poppy results. I sorta like it, it's different.