HOME | DD

Description
Another attempt at using img2img on an image I've used previously. This came out a lot better (far more faithful to the original) and while a new model might've helped somewhat (since the Furry model actually has tags for elves), I think the success of this attempt was down to my new approach to img2img-ing.

Until now, I had this unhelpful restraint when it came to intervening in img2img generations - remain faithful to the starting images and its products, relying solely on what the AI produces for the process. It meant that I'd have to run dozens of extra attempts to get something usable, just sitting and rerolling every time the AI added a fat roll where it shouldn't be, or placed the navel in an odd position. I realised that approach was stupid. Instead, this time I'd iterate with the AI until I found something good, while making my own edits to help things along - stuff like introducing fat rolls where they should be or where they were removed, nudging elements that I know the AI is going to misinterpret, and other things such as that. Heck, I even did a bit of regular image editing on the final product to get rid of an artifact/blemish on the final result that I hadn't noticed at the time.

If you're wondering why the face has seen so much of an improvement, that's thanks to me learning a bit more about inpainting. Using the high resolution mode, you can effectively target areas to re-make at a higher resolution to great effect, as you can see here. I used the technique to get an actually good-looking face (there were consistently good ones to choose from, and I liked this one with the big hair) as well as increase the detail of the bra area, as it was consistently lacking in good detail work due to the original image (used inpainting and altering the prompt to include "highly detailed green bra" with certain weights until I got something shaded nicely). The image probably could've done with another few passes to iron out the background - I hadn't noticed that the inpainting of the head had affected it!

The whole iteration and editing does leave me at a bit of a dilemma, however. While curated img2img does definitely yield better results (well, results closer to what I desire), it does start to stray away from one of the things I wanted for this page. The more iteration and editing that goes on, the less and less meaningful all the prompt and setting information becomes. There's no way that anyone who hopes to follow along will be able to use this information to yield anything resembling this - there's been too much iteration, as well as tweaks to the prompt during the process for what's there to mean anything. It's a shame, because the ability for others to be able to follow along and replicate everything I do is something I hold fairly highly... Though I'm not sure how many people really care about it Xb

I think in the future I may begin cutting out most of the AI settings information on img2img generations as they become less useful. To make up for it, I'll try to talk more about the process of what I did to create the image.

Base image is cropped from: EFS 49

Disclaimer: img2img is an iterative process, usually involving feeding output images back into the model to come closer to the ultimate image. While the starting image and generation data are provided, you shouldn't expect the base image to directly yield the final images seen here.

Prompt: humanoid female elf, solo, blonde hair, green bra, huge breasts, morbidly obese, hyper belly, belly overhang, (fat rolls:1.3), art by artgerm and thomas kinkade, fantasy tavern in background, highly detailed green bra, cleavage
Negative prompt: deformed, (legs:1.5), (thighs:1.5), feet, navel
Seed: 4073203457
Other: Steps: 20, Sampler: Euler a, CFG scale: 8, Denoising strength: 0.7, Mask blur: 4 Size: 512x512
Model: Furry [323f8dd8]

Description
Another attempt at using img2img on an image I've used previously. This came out a lot better (far more faithful to the original) and while a new model might've helped somewhat (since the Furry model actually has tags for elves), I think the success of this attempt was down to my new approach to img2img-ing.

Until now, I had this unhelpful restraint when it came to intervening in img2img generations - remain faithful to the starting images and its products, relying solely on what the AI produces for the process. It meant that I'd have to run dozens of extra attempts to get something usable, just sitting and rerolling every time the AI added a fat roll where it shouldn't be, or placed the navel in an odd position. I realised that approach was stupid. Instead, this time I'd iterate with the AI until I found something good, while making my own edits to help things along - stuff like introducing fat rolls where they should be or where they were removed, nudging elements that I know the AI is going to misinterpret, and other things such as that. Heck, I even did a bit of regular image editing on the final product to get rid of an artifact/blemish on the final result that I hadn't noticed at the time.

If you're wondering why the face has seen so much of an improvement, that's thanks to me learning a bit more about inpainting. Using the high resolution mode, you can effectively target areas to re-make at a higher resolution to great effect, as you can see here. I used the technique to get an actually good-looking face (there were consistently good ones to choose from, and I liked this one with the big hair) as well as increase the detail of the bra area, as it was consistently lacking in good detail work due to the original image (used inpainting and altering the prompt to include "highly detailed green bra" with certain weights until I got something shaded nicely). The image probably could've done with another few passes to iron out the background - I hadn't noticed that the inpainting of the head had affected it!

The whole iteration and editing does leave me at a bit of a dilemma, however. While curated img2img does definitely yield better results (well, results closer to what I desire), it does start to stray away from one of the things I wanted for this page. The more iteration and editing that goes on, the less and less meaningful all the prompt and setting information becomes. There's no way that anyone who hopes to follow along will be able to use this information to yield anything resembling this - there's been too much iteration, as well as tweaks to the prompt during the process for what's there to mean anything. It's a shame, because the ability for others to be able to follow along and replicate everything I do is something I hold fairly highly... Though I'm not sure how many people really care about it Xb

I think in the future I may begin cutting out most of the AI settings information on img2img generations as they become less useful. To make up for it, I'll try to talk more about the process of what I did to create the image.

Base image is cropped from: EFS 49

Disclaimer: img2img is an iterative process, usually involving feeding output images back into the model to come closer to the ultimate image. While the starting image and generation data are provided, you shouldn't expect the base image to directly yield the final images seen here.

Prompt: humanoid female elf, solo, blonde hair, green bra, huge breasts, morbidly obese, hyper belly, belly overhang, (fat rolls:1.3), art by artgerm and thomas kinkade, fantasy tavern in background, highly detailed green bra, cleavage
Negative prompt: deformed, (legs:1.5), (thighs:1.5), feet, navel
Seed: 4073203457
Other: Steps: 20, Sampler: Euler a, CFG scale: 8, Denoising strength: 0.7, Mask blur: 4 Size: 512x512
Model: Furry [323f8dd8]