WARNING: Highly experimental.
V2 Z-Image Turbo: 1-2 weight strength for more glitch. 0.25-1 for more subtle look
Add the following prompt after your subject to improve likeness of scrambled cable tv.
"A distorted, abstract television screen resembling scrambled analog cable channels from the 1990s, image should feature wavy horizontal lines, warped color bands, static noise, shifting rainbow streaks, ghosted shadows of shapes, and intermittent flickering patterns. chaotic, glitchy, analog interference of pay-per-view channels before decoding. The aesthetic should feel retro, analog, low-resolution, filled with signal noise, chromatic aberration, and VHS-like texture,"
This prompt was made using ChatGPT and improves the look. It might work on the Flux versions, but I haven't tried it.
V2:
This is trained on frames of a scrambled 90s cable tv channel. The source footage was deinterlaced, so the renders can look blurry. Combining it with a sharpen lora or another glitch lora can add some details. Use "color" for color generations that lean heavily into the analog blue and purple chromatic aberration. If there is a specific color you want for something, include that in the prompt (ex: red shirt) Just using "scr4mb13dtv, glitch" without any subjects will produce the abstract analog glitching of the signal being out of sync and scrambled. The images used for training were SFW.
V1:
This is trained on frames of VHS recordings from scrambled analog cable tv. Some trigger words are scrmblcbletv, glitch, scrambled, color, interlaced, analog. Adding distortion, distorted, green skin, chromatic aberration, vhs style, etc can get interesting results even though it wasn't trained on those words. Under 1 weight brings out subjects better and over 1 gets more glitched out. There was no nsfw content in the training data, so it won't produce any nsfw content unless you combine it with a nsfw lora or checkpoint.