How to refine your prompt in Stable Diffusion

Refine is a Telegram Bot command that lets you quickly re-prompt a rendered image. Use the refine command to make a quick change without having to use copy/paste or type the whole prompt again.


Render an image like you normally would. Wait for the results to arrive, and think about how you’d like to improve or change the next set of images. To then refine that prompt:

Step 1:  Select any of those images (doesn’t matter which)
Step 2: Hit the REPLY button,
Step 3: Type:


The flow looks like this:

The Refine button will appear, as shown above. Click the link or the button and a pop-up will let you refine your prompt.

Use Refine when you want to start over, but work from your prompt.


It must be a rendered image with a clean prompt.

This won’t work with pasted images or, in some cases, forwarded messages. A malformed prompt may also cause refine to fail, in the case of too many mixed parenthesis or square brackets that were not properly terminated.


Why can’t I just edit the text?

Editing what is on the screen doesn’t change the fact that the train has already left the station.

In Telegram, a prompt is sent like a chat message, so it’s gone from the text area. There’s nothing we can do (at this time) to move that text back down, as it is sent off to the GPU servers for rendering.

Refine is thus used when we have waited for the results, the image has arrived, and we want to make a change. Instead of scrolling back up to find the prompt and copying/pasting, just use refine instead.

Refine sends you to our new weights editor which makes it easy to visualize concepts, recipes, and other powerful settings.


If you change your mind, you can cancel the queue before it arrives with a different cancel command. You can’t refine after a cancel, because refine requires an image as input.


If you’d like to continue from a specific image of the set instead of a general reprompt, first grab the seed value of that image by inspecting it.  You can then paste that number into refine. Tap the image to see the number, and then add /seed:###### to the positive area of the refine box.


Refine is literally editing your own text and choosing the AI models that you want to run. Thus, the image is not determined yet, and no image was used as input.  You are starting from scratch with Refine.

Remix is the more powerful but destructive version, where every pixel is changed to execute the new instructions using the AI model in the original prompt, or the new model you specify. So the old image is still in play.

Edit is its own AI model, which allows you to ponder changes to an image using natural language. The image is receiving new instructions and using the old image as an input.  Edit is best at making minor changes, like asking “What if her hair was blue?”  So the old image is used to determine the result.  Because it uses its own AI model, this model is hard to recommend in some cases (you can’t target a concept)

We apologize in advance for the similar names, but these features were rolled out at very different times. In the future, we may rename them something much more distinctive, like Peter, Mary, and Obama.

Jokes aside, we’d love your input on this. Chat with us in VIP or suggest a feature directly to our roadmap.