"I forgot what I wanted because I got it. And, gosh, I wish I could want again."

"I forgot what I wanted because I got it. And, gosh, I wish I could want again."

Here's my entry for @ClaireSilver12's contest "Back to School." #ClaireAIContest.

I focused on conveying an emotion and capturing the essence of an era, mostly an idealized one with visual codes borrowed from "college" TV shows and movies I enjoy. I have written a complete wrap-up of my experience, but TL;DR:

- Masking techniques
- Depth maps and camera movements
- @StabilityAI's technology
- ComfyUI
- @midjourney
- @LeonardoAi_
- @Magnific_AI
- @pika_labs
- @Blender
- @3DAIStudio
- @suno_ai_
- @elevenlabsio

I spent a lot of time experimenting. The hardest part was rejecting many shots because they served more as technical demonstrations rather than contributing to the narrative. I discussed this with Dina (@Ethereal_Gwirl) and expressed that I didn't want my piece to look like a demo reel.

That being said, each shot includes a technical twist. For example, in the first shot, I initially used birds generated in Blender. However, they lacked grit and were too perfect, so I ran them through AnimatedDiff and then brought them back into the shot, giving them a more vintage look.

In this piece, I have worked extensively on masking techniques. I used the methods demonstrated by @mickmumpitz in his "AI rendering V2" video (link below).

Nearly every shot incorporates depth maps to create dynamic camera movements. I utilized @LeonardoAi_ and displacement modifiers in Blender.

All the images were generated in ComfyUI. Sometimes, I would generate images in Midjourney and use ControlNet to replicate the composition using my Loras (I trained one specifically for this contest). Each image was improved using @Magnific_AI (I always started with it like you would start your day with a cup of coffee).

The girl in the first shot was animated using @pika_labs. Other shots relied on SVD, which does an awesome job (though you have to interpolate like crazy).

The cafeteria shot received special treatment: I created a desk using @3DAIStudio, roughly composed the shot in Blender, rendered a "freestyle only" version of it, and used ControlNet's LineArt processor to get the final render in ComfyUI. However, when I couldn't achieve the exact desired result, I turned to Midjourney, using my ComfyUI image as a reference with --iw 3. And it worked perfectly.

The voiceover was created using @elevenlabsio. ChatGPT helped ensure my text was error-free; otherwise, I didn't solicit its assistance.

I composed the music in Ableton Live. Just as I was finishing, I received a message from @suno_ai_ offering early access to their new features, including the ability to create a song from any sound. I used my track as input, then downloaded the track Suno provided and blended it into my original composition. The resulting texture is quite impressive.

At the end of the process, I used After Effects and Da Vinci Resolve for post-processing and editing, and @topazlabs for upscaling.

It was really fun experimenting with new techniques and taking the time to try new things. Thank you for the opportunity, Claire.