# How to get intermeidate output images
**URL:**
**Author:**  [@dkackman](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Fu%2Fdkackman)%5C
**Post date:** [January 7, 2023, 11:49pm UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fhow-to-get-intermeidate-output-images%2F29144%2F1 "2023-01-07T23:49:56Z")
Is it possible to get the images at each denoising step via the Diffusers library? I am sure I’ve seen it done but can’t find where or how.
---
**Author:**  [@pcuenq](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Fu%2Fpcuenq)%5C
**Post date:** [January 8, 2023, 11:34am UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fhow-to-get-intermeidate-output-images%2F29144%2F2 "2023-01-08T11:34:39Z")
Hi @dkackman!
You might want to look at the _callback mechanism_, which sends intermediate latents to a function you specify. You could then decode the latents in that function and visualize them as you need.
[This notebook includes a section about callbacks](https://github.com/fastai/diffusion-nbs/blob/master/stable_diffusion.ipynb) that demonstrates how to use that feature.
Good luck!
---
**Author:**  [@dkackman](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Fu%2Fdkackman)%5C
**Post date:** [January 8, 2023, 5:53pm UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fhow-to-get-intermeidate-output-images%2F29144%2F3 "2023-01-08T17:53:34Z")
Oh perfect. I was unclear on how to transform the latents into an image but this exactly what iI was looking for.
```python
vae = pipe.vae
images = []
def latents_callback(i, t, latents):
latents = 1 / 0.18215 * latents
image = vae.decode(latents).sample[0]
image = (image / 2 + 0.5).clamp(0, 1)
image = image.cpu().permute(1, 2, 0).numpy()
images.extend(pipe.numpy_to_pil(image))
prompt = "Portrait painting of Jeremy Howard looking happy."
torch.manual_seed(9000)
final_image = pipe(prompt, callback=latents_callback, callback_steps=12).images[0]
images.append(final_image)
image_grid(images, rows=1, cols=len(images))
```
---
**Author:**  [@venkatesh-thiru](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Fu%2Fvenkatesh-thiru)%5C
**Post date:** [March 17, 2025, 5:55pm UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fhow-to-get-intermeidate-output-images%2F29144%2F4 "2025-03-17T17:55:44Z")
Whats with the scaling in `latents = 1 / 0.18215 * latents`? is it a constant for every VAE? can I still apply the same callback for SD3.5?
---
**Author:**  [@John6666](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Fu%2FJohn6666)%5C
**Post date:** [March 18, 2025, 6:02am UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fhow-to-get-intermeidate-output-images%2F29144%2F5 "2025-03-18T06:02:32Z")
I think the same method can be used for the Diffusers pipeline.
### Pipeline callbacks
> **[Pipeline callbacks](https://huggingface.co/docs/diffusers/using-diffusers/callback)**
>
> We’re on a journey to advance and democratize artificial intelligence through open source and open science.
### Explanation of the 0.18215 factor in textual\_inversion?
>
>
> https://github.com/huggingface/diffusers/blob/b2b3b1a8ab83b020ecaf32f45de3ef2364…4331cf/examples/textual\_inversion/textual\_inversion.py#L501
>
> Hi, just a small question about the quoted script above which is bothering me: where does this \`0.18215\` number come from? What computation is being done? Is it from some paper? I have seen the same factor elsewhere, too, without explanation. Any guidance would be very helpful, thanks!