Logo_Iteration / README.md
Yankkee's picture
Upload 3 files
ea664ce verified
|
Raw
History Blame Contribute Delete
2.45 kB
---
title: Mass Iteration Studio
emoji: πŸŒ€
colorFrom: indigo
colorTo: pink
sdk: gradio
sdk_version: 6.5.1
app_file: app.py
pinned: false
license: mit
short_description: One image in, N PNG variants out
---
# Mass Iteration Studio
Upload one image, get N variants back as PNG. Color, style and form are sampled
per variant from editable prompt pools. Requires **ZeroGPU** hardware.
> Set `sdk_version` above to the value Hugging Face wrote into the README it
> generated for your Space.
## Setup
1. New Space β†’ SDK **Gradio** β†’ Settings β†’ Hardware β†’ **ZeroGPU**.
2. Upload `app.py`, `requirements.txt`, `README.md`.
3. First run downloads ~7 GB of weights and takes several minutes. Every run
after that starts instantly β€” the pipeline stays in memory.
## How the run is structured
A 20-image run is not one GPU call. Variants are grouped by denoising strength
and rendered in batches, each batch its own `@spaces.GPU(duration=75)` call.
That keeps every call well under the platform ceiling and lets results stream
into the gallery as they finish, instead of appearing all at once at the end.
## Models
| Option | Time per image | 20 variants | License |
|---|---|---|---|
| SDXL-Lightning 4-step (default) | ~0.8 s | ~16 s | OpenRAIL++ base, Apache 2.0 LoRA |
| SDXL-Turbo 2-step | ~0.45 s | ~9 s | non-commercial |
| SD-Turbo 2-step | ~0.2 s | ~4 s | non-commercial |
Pro quota is 1500 GPU-seconds per day, so the default model gives roughly 90
runs of 20 variants before the window resets.
## The reinvention slider
One control decides everything about how far variants drift:
- **0.2–0.35** β€” recolor and restyle, subject untouched
- **0.4–0.55** β€” forms shift, proportions change, subject still recognizable
- **0.65–0.85** β€” only the rough composition survives
Each variant draws its own value between your minimum and maximum, snapped to
six discrete levels so same-level variants can share a batched GPU call.
## Prompt pools
Three editable text boxes, one option per line. Sampling is either a random mix
or a grid sweep that walks every combination in order β€” useful when you want
systematic coverage rather than a lucky draw.
Every run writes `manifest.csv` into the ZIP with the prompt, strength and seed
behind each file. To rebuild a single variant exactly, set the base seed and
turn off randomization.
## Output
PNG only. Individual downloads from the gallery, or the full run as a ZIP.