The main difference between `img2img` and `prompt2img` is the starting point. While `prompt2img` always starts with pure
gaussian noise and progressively refines it over the requested number of steps, `img2img` skips some of these earlier steps
(how many it skips is indirectly controlled by the `--strength` parameter), and uses instead your initial image mixed with gaussian noise as the starting image.
**Let's start** by thinking about vanilla `prompt2img`, just generating an image from a prompt. If the step count is 10, then the "latent space" (Stable Diffusion's internal representation of the image) for the prompt "fire" with seed `1592514025` develops something like this:
Put simply: starting from a frame of fuzz/static, SD finds details in each frame that it thinks look like "fire" and brings them a little bit more into focus, gradually scrubbing out the fuzz until a clear image remains.
**When you use `img2img`** some of the earlier steps are cut, and instead an initial image of your choice is used. But because of how the maths behind Stable Diffusion works, this image needs to be mixed with just the right amount of noise (fuzz/static) for where it is being inserted. This is where the strength parameter comes in. Depending on the set strength, your image will be inserted into the sequence at the appropriate point, with just the right amount of noise.
### A concrete example
Say I want SD to draw a fire based on this hand-drawn image:
Let's only do 10 steps, to make it easier to see what's happening. If strength is `0.7`, this is what the internal steps the algorithm has to take will look like:
With strength `0.4`, the steps look more like this:
Notice how much more fuzzy the starting image is for strength `0.7` compared to `0.4`, and notice also how much longer the sequence is with `0.7`:
| | strength = 0.7 | strength = 0.4 |
| -- | -- | -- |
| initial image that SD sees |  |  |
Both of the outputs look kind of like what I was thinking of. With the strength higher, my input becomes more vague, *and* Stable Diffusion has more steps to refine its output. But it's not really making what I want, which is a picture of cheery open fire. With the strength lower, my input is more clear, *but* Stable Diffusion has less chance to refine itself, so the result ends up inheriting all the problems of my bad drawing.
If you want to try this out yourself, all of these are using a seed of `1592514025` with a width/height of `384`, step count `10`, the default sampler (`k_lms`), and the single-word prompt `fire`:
The code for rendering intermediates is on my (damian0815's) branch [document-img2img](https://github.com/damian0815/InvokeAI/tree/document-img2img) - run `dream.py` and check your `outputs/img-samples/intermediates` folder while generating an image.
The code for rendering intermediates is on my (damian0815's) branch [document-img2img](https://github.com/damian0815/InvokeAI/tree/document-img2img) - run `invoke.py` and check your `outputs/img-samples/intermediates` folder while generating an image.
After putting this guide together I was curious to see how the difference would be if I increased the step count to compensate, so that SD could have the same amount of steps to develop the image regardless of the strength. So I ran the generation again using the same seed, but this time adapting the step count to give each generation 20 steps.
Here's strength `0.4` (note step count `50`, which is `20 ÷ 0.4` to make sure SD does `20` steps from my image):
In both cases the image is nice and clean and "finished", but because at strength `0.7` Stable Diffusion has been give so much more freedom to improve on my badly-drawn flames, they've come out looking much better. You can really see the difference when looking at the latent steps. There's more noise on the first image with strength `0.7`:
than there is for strength `0.4`:
and that extra noise gives the algorithm more choices when it is evaluating how to denoise any particular pixel in the image.
Unfortunately, it seems that `img2img` is very sensitive to the step count. Here's strength `0.7` with a step count of `29` (SD did 19 steps from my image):
By comparing the latents we can sort of see that something got interpreted differently enough on the third or fourth step to lead to a rather different interpretation of the flames.
This is the result of a difference in the de-noising "schedule" - basically the noise has to be cleaned by a certain degree each step or the model won't "converge" on the image properly (see https://huggingface.co/blog/stable_diffusion for more about that). A different step count means a different schedule, which means things get interpreted slightly differently at every step.
