ComfyUI Image & Video Studies
Local experiments in directing images and video through ComfyUI’s node-based workflows.
- My contribution
- Independent research, workflow experimentation, image direction and iterative inpainting. Later workflow development with Codex assistance.
- Project tools
- Stable Diffusion · Inpainting · ComfyUI · OpenAI Codex · MiniMax H3 / Director
ComfyUI workflows
My ComfyUI exploration began in 2023, when my sister asked whether I could turn her drawings into 3D. Building them in 3D would take time, so I started investigating what I could achieve through local image generation with Stable Diffusion.
Abyssal Elegance
For Abyssal Elegance and Cosmic Bloom, I used ControlNet, depth maps and repeated inpainting to work toward the images I wanted. It felt like painting with AI: directing, adjusting and trying again until the result looked right to me. My sister’s drawings provided the starting compositions, and my contribution was interpreting them through that process.
Cosmic Bloom
Video studies
I researched and developed those early studies independently. More recently, I have used Codex to help prepare and troubleshoot workflows as I explore video generation in ComfyUI.
The selected MiniMax H3 tests use familiar characters in comic situations. One draws on the bench and box of chocolates from Forrest Gump; another places Spider-Man and Pikachu at a music festival. I wanted to see how the model interpreted characters, film references and a scene I had in mind, while exploring what I could achieve on my RTX 3090.
The videos did not follow every part of my intention, but the results encouraged me to keep experimenting. I see them as material to refine through further workflow adjustments and editing. Across both image and video, my interest is in learning how to direct these tools toward a particular idea.