*This is part 1 of a 2 part series. See the semi-realistic illustrations at part 2: https://civitai.com/posts/5012271

An Ode to img2img

This was my first time experimenting with img2img, and while I've ignored this new feature until now, I now see it as an indispensable tool for helping realize artistic vision and intent. From my initial impressions, it felt like it did a better job at denoising and adding details than the more expensive upscale + hi-res fix. But I'm beginning to think that the real power in img2img lies in its ability to steer image generation through changing of the prompt, parameters, LoRAs, and checkpoint mid generation. For instance, I toggled checkpoints between anime to realistic each time I iterated with img2img on a particular generation to encourage cross-pollination and to compensate for weaknesses in prompt adherence and aesthetics across checkpoints. It's still not at the level of control I want for generation, but adding img2img is a huge improvement over the feeling of rolling the dice.

Approximate workflow (all done on Civitai):

1. Draft mode (crank out 4 low-cost images)

2. Pick most interesting image based on composition, etc.

3. img2img to remove noise, add detail, and guide generation

4. Repeat step 2 until image is cleaned up

5. Upscale 1.5x + hi-res fix

6. img2img one last time