A big problem I’ve been having is spending too long on 1 picture, so the challenge is to come up with a workflow where you spend no more than a few minutes on each picture.
My workflow was the following:
1. Great model, lots of negative textual inversions
2. Simple prompt where I just change the description of the subject
3. Quickly generate a dozen images
4. Pick the best one and send it to img2img for 1.5x increase and play around with CFG and denoise values a few times till I’m happy
5. Don’t worry about minor error, if you really like a picture you can go back to it another time and fix it up
6. If the process is taking more than 5 minutes move on to a different image/prompt
Here was my prompts roughly:
> {subject description}, kkw-ph1, soccer, athletic, sexy, lesbian, soccer ball, (topless:1.3), soccer shorts, kicking ball, realistic proportions, (big boobs:0.6)
> Negative prompt: CyberRealistic_Negative-neg, FastNegativeV2, kkw-Extreme-Neg, black and white, monochrome, camera, blurry, bokeh, super model
> logoSteps: 28, Sampler: DPM++ 2M SDE Karras, CFG scale: 7.5, Seed: 847271570, Size: 768×1152, Model hash: 7f6146b8a9, Model: edgeOfRealism_eorV20Fp16BakedVAE, Denoising strength: 0.75, Token merging ratio: 0.5, Version: v1.3.2
