Create a training job, or pick one on the left.

Download LoRA

Your LoRA is ready to use

Put this tag anywhere in your prompt, in the bot or on the website:

  • The long code is the LoRA's hash. Its first ten characters work too:
  • The number at the end is the strength. 1 is full strength; try 0.6 to 0.8 if the result looks overcooked.
  • It is private: it does not appear in the LoRA list, and it has no name, only this tag. Anyone you give the tag to can use it too, so share it only with people you want to have it.
  • "Download LoRA" above gives you the file itself (.safetensors) for use elsewhere.
Log 

                        

                    


                    
                    

                        

                            
Settings

                            
                        


                        

                            
                            


                            
                            

                            
                            
                            

                            
                            
                            
                            
                            
                        


                        
Adapter

                        

                            
                            
                            
                            
Options
                                

                                    
                                    
                                

                            

                            
Trained layers
                                

                                    
                                    
                                    
                                    
                                    
                                    
                                

                            

                        


                        
Optimizer

                        

                            
                            
                            

                            
                            
                            
                            
                            
                            
                        


                        
Noise and captions

                        

                            
                            
                            
                            
                            
                            
                            
                            
                            
                            
                            
Options
                                

                                    
                                    
                                    
                                

                            

                            
                        


                        
Preview images

                        
Rendered by the burrow during training, with and without the LoRA. Leave empty for none.
                            Write each prompt as you would write one to generate a picture: the trigger word, then what
                            the picture should show (the trigger word alone says nothing about the scene). Every epoch
                            is rendered with the same seed, so the epochs can be compared.
                            Each preview has its own prompt, negative prompt, size, steps, CFG scale and seed. Previews are
                            a plain single pass (no hires step), so a normal generation with the finished LoRA usually
                            looks better than they do.

                        

                        

                            
                            
                            
Options
                                

                                    
                                

                            

                        


                        

                            Advanced parameters (JSON, merged over the request)
                            
Any parameter the burrow accepts, as a JSON object. Example:
                                {"noise": {"multires_iterations": 6}, "bucket": {"min": 768, "max": 1344}}

                            
                        

                    


                    
                    

                        

                            
Dataset 

                            

                                
                                
                            

                        

                        

                            Drop images or a .zip of them here, as many as you like at once. Caption files that come
                            with them are matched by name: fox.png takes fox.txt (tags) and
                            fox.natural.txt (a sentence). A zip is opened in your browser, and only the images
                            and caption files in it are sent.
                            
                        


                        

                            
                            
Trigger word
                                

                                    
                                

                            

                            

                                
                                
                            

                        

                        
The tagger is the one the bot already uses to describe images. Its tags go in
                            the tag caption of each image, where you can still edit them; the natural-language caption is
                            never touched. With "Fold what the images share" on (and a trigger word set), tags found in at
                            least four of every five captions (blue fur, goggles on every picture
                            of one character) are left out of training, so the trigger word alone brings them; what differs
                            between images stays something you ask for in the prompt. Your captions here are not changed,
                            and the job's log lists what was folded.


                        

                            

                            

                                
                                
                                
                            

                            

                        

                        

                    


                

            

        

    


    
    

        

        

            
Add from my generations

            

                
Newest first. Click the pictures you want; they are added to this job's dataset, not copied.

                

                

                    
                    
                

            

            

                
                
            

        

    


    
    

        

        

            
Make this LoRA public

            

                

                
    
                    
  • It is public for good. Once it is approved it cannot be removed or made private again.
    • 
                    
  • Anyone can use it, by its name, in the bot and on the website.
    • 
                    
  • You give it a title, a description, its trigger words and at least one example picture on the
                        next page. Staff look at it before it goes public, and may turn it down.
    • 
                

                

                
                

            

            

                
                
            

        

    


    
    

        

        Preview