RandyCamp Lo-Fi
Glitch Lab
I wanted a fast way to make a lot of rough visual material for RandyCamp without treating cloud GPUs as the default answer. This page documents the system I built around a local 8GB RTX card, the decisions that kept it stable, and the 500 stills it produced.
A generator for quantity, not a single perfect image.
RandyCamp visuals need volume. A live deck gets boring when it depends on a few hero clips. The useful question was whether a small local GPU could generate enough variation to feed Clipcycle: still cards, ghost overlays, background texture, and raw material for later video work.
The answer was yes, if the system respected the hardware. I used lightweight diffusion models, fp16, 512px images, and a resumable img2img batch script. The work stayed local, the outputs landed directly in the RandyCamp drop folder, and the evaluation page made the results easy to inspect on a phone.
How the system was built.
Start with a small card, not a cloud bill
The goal was not to chase the largest model. The goal was to find out how much useful visual material a local RTX 2080 SUPER with 8GB of VRAM could produce without renting a GPU, waiting in a queue, or turning a live visual system into a cloud dependency.
Pick models that fit the machine
The practical lane was SD 1.5-class image generation at 512 by 512, fp16, batch size 1. That kept the work inside the card instead of asking an 8GB machine to behave like a modern video workstation.
Build a test set before the big run
The first smoke test generated 50 stills: ten prompts across five models. That was enough to see which models had a visual identity, which prompts failed, and whether the system could run repeatedly without exhausting VRAM.
Turn the smoke test into a variation machine
The second pass used each of the 50 source stills as an img2img input. Every source went through every model twice: one RGB melt pass and one signal decay pass. That produced 500 RandyCamp-ready stills for Clipcycle.
The 8GB card handled the job because the job was sized correctly.
The RTX 2080 SUPER did not behave like a modern high-VRAM generation rig. It behaved like a good constraint. At 512px, batch size 1, and SD 1.5-class checkpoints, the run was stable and fast enough to be useful. The full batch finished after a 20-image proof run and a 480-image continuation run.
The combined GPU generation time was 12m 32s. The full run reported 500 unique PNG outputs, no low-contrast failures, and no dimension mismatches. During the run the card stayed under load rather than thrashing. That matters more than benchmark theater: the system produced a usable deck.
What this card should and should not be asked to do.
- 512 by 512 was the honest resolution for this run. It kept the card stable and fast.
- Batch size stayed at 1. Bigger batches would have raised the chance of VRAM failure for no useful gain in this workflow.
- The models were SD 1.5-era models. SDXL, Flux, and video generation are different hardware problems.
- The public gallery uses WebP copies. The raw PNG deck is larger and stays in the local RandyCamp folder.
- Signal decay is intentionally quieter than RGB melt. It works as a ghost layer, but a louder third glitch mode would be useful later.
Five lanes, one hundred stills each.

DreamShaper 8
best balanced lane
100 stills, 1.461s average
Inkpunk Diffusion
most graphic lane
100 stills, 1.466s average
Analog Diffusion
lo-fi photo texture lane
100 stills, 1.466s average
Counterfeit V2.5
anime background lane
100 stills, 1.478s average
Stable Diffusion 1.5 baseline
raw control lane
100 stills, 1.463s averageAll 500 generated stills.
Each card opens the larger WebP version. The original PNG deck is kept locally for Clipcycle.
DreamShaper 8100 stills, best balanced lane
The strongest default. It held scenes together while adding enough chroma damage to feel like a live visual source.
Inkpunk Diffusion100 stills, most graphic lane
The most distinctive lane. Hard outlines, poster energy, and blacklight comic texture made it the best punch card.
Analog Diffusion100 stills, lo-fi photo texture lane
The softest lane. It traded sharpness for film haze, photocopy texture, and old-camera atmosphere.
Counterfeit V2.5100 stills, anime background lane
The cleanest stylized lane. It leaned toward anime backgrounds and calmer interiors.
Stable Diffusion 1.5 baseline100 stills, raw control lane
The control lane. It was rawer and less composed, but useful because roughness can be an asset in a glitch deck.