Prompts weren’t enough
I started my journey into AI image generation in February 2025. I was amazed that I could provide a list of objects, scenery, and poses to a computer program, and it would create an image containing most of those things, more or less the way I described them. But as I continued to explore the medium, I became frustrated by the lack of depth and atmosphere. As a ComfyUI user, I experimented with hundreds of LoRAs and checkpoints, and I endlessly tweaked prompts, weights, and wording. I knew prompts alone weren’t giving me the atmosphere I wanted, but I had no idea how to communicate that to the model. Then, almost by accident, I made a discovery that changed not only how I thought about the process, but how I think the model interprets it.
One of my original ComfyUI workflows included both text-to-image (T2I) and image-to-image (I2I) paths, allowing me to generate an image and then immediately refine or upscale it without changing workflows. The process was straightforward: write a prompt → generate an image → copy it to the input folder → load it into a Load Image node → send it through a KSampler with a denoise value between 0.3 and 0.5 → enjoy. I even built a collection of switches and groups to make moving between the two modes almost effortless.
And it worked beautifully.
Until one day, I forgot to switch back to T2I mode before generating an entirely different image from a completely new prompt. That one little mistake opened up a world of possibilities and significantly changed how I think about the images I create.
In the next installment, I’ll show you exactly what happened when I accidentally left Image-to-Image mode enabled, why the result surprised me, and how it led me down a completely different path.
This isn’t a tutorial or a secret workflow. It’s simply the story of one happy accident and the ideas it inspired.