[Seizure Warning] Bad Apple, but Played Through AI-generated Paintings

Like it says in the title. For the record, this is an HD reupload. The original 480p video proved to be too crunchy: I spent 13ish hours in total generating the images and... less than 15 minutes editing it and rendering it. I DO NOT condone the use of AI image generation for personal gain, I suggest you draw inspiration from it or do stupid things like this instead. Here, have some downloads so you can see the individual frames in all their glory: SD: #4JdVV8a6Q35erOLp32sAw2QNdN1fi0MTyWKBSSjE8T4 HD: #x3aoAAAmK7U8toXdaQqTwNlmWzzNdmFG7gJbJzJ_EnQ Just the frames: #5iG1bgvnqendpfYo6k_-ag Here’s all the things you need to make this yourself! ========================================== 0. a fairly decent computer (you need a GPU with 4 GB of VRAM at a minimum to run SD) 1. a WebUI to make this easier: AUTOMATIC1111’s SD webui: The Controlnet extension for the webui: 2. a couple models: Realistic Vision (put in the webui’s ’models/Stable-diffusion’ folder): QR Code Monster (download the .safetensor file, put in ’models/Controlnet’ or ’extensions/sd-webui-controlnet/models’): 3. some Textual Inversion models (put in webui’s ’embeddings’ folder) bad-artist-anime: bad_prompt_version2: bad-hands-5: ng_deepnegative_v1_75t: 4. A directory with frames of whatever you want to AI-ify. You can take any mp4 you have and split it to a series of numbered frames using FFMpeg (or just download a repository of frames if you want to do bad apple, which I did for this video). 5. the most important step, proper settings: First, the Positive Prompt (you can modify the plain sentence after the BREAK and newline for a different style, I just used the prompt for that popular spiral image): (masterpiece:1.4), (best quality), (detailed), BREAK painting of a medieval village scene with busy streets and castle in the distance Then, the Negative Prompt (this is to discourage the AI from drawing things it finds in the contours that deviate from the prompt): (worst quality, poor details:1.4), lowres, (artist name, signature, watermark:1.4), bad-artist-anime, bad_prompt_version2, bad-hands-5, ng_deepnegative_v1_75t, weird shadows Whole bunch of miscellaneous SD settings: Steps: 20, Noise Sampler: Euler a, CFG scale: 7 Then some Controlnet settings: Weight: 1.5, Guidance Start: 0, Guidance End: 1, Pixel Perfect: True, Control Mode: Balanced Tick the ’Increment seed after each controlnet batch iteration’ setting in the Webui’s settings/Controlnet tab if you want the frames to look distinct between each other like this video. 6. Lastly, some general instructions for setup: Select the ’realistic-vision’ model in the main SD settings, get the QRCode monster model in the controlnet settings. I highly recommend doing single frames and tweaking settings until you like what you see. When you’re ready to AI-ify the frames, switch Controlnet to ’Batch’ and point it toward the directory full of your video’s frames, then click ’generate’ in the upper-right of the webui. Once it’s done generating, go to the webui’s outputs/[date] folder (and maybe copy them to someplace else), rename the frames so they’re in increasing numerical order, then open the PNG sequence in your favorite video editor and sync up the audio. Be sure to comment if you have any questions with the process above (or, y’know, just search online for it) Have fun! May the GPU’s of heretics suffer.