Notebook links

README.md

@@ -50,12 +50,14 @@ base.load_lora_weights("minimaxir/sdxl-wrong-lora")
 _ = base.to("cuda")
 ```
 
-During inference, use `wrong` as the negative prompt.
+During image generation, use `wrong` as the negative prompt.
 
 ## Examples
 
 **Left image** is the base model output (no LoRA) + refiner, **right image** is base (w/ LoRA) + refiner. Both generations use the same seed.
 
+I have also [released a Colab Notebook](https://colab.research.google.com/github/minimaxir/sdxl-experiments/blob/main/sdxl_wrong_comparison.ipynb) to generate these kinds of images, although the seeds listed will not give the same results since they were generated on a different GPU/CUDA than the Colab notebook.
+
 `realistic human Shrek blogging at a computer workstation, hyperrealistic award-winning photo for vanity fair` (cfg = 13, seed = 56583700)
 
 ![](img/example1.webp)
@@ -80,8 +82,11 @@ During inference, use `wrong` as the negative prompt.
 
 The methodology and motivation for creating this LoRA is similar to my [wrong SD 2.0 textual inversion embedding](https://huggingface.co/minimaxir/wrong_embedding_sd_2_0) by training on a balanced variety of undesirable outputs, except trained as a LoRA since textual inversion with SDXL is complicated. The base images were generated from SDXL itself, with some prompt weighting to emphasize undesirable attributes for test images.
 
+You can see the code to generate the wrong images [in this Jupyter Notebook](https://github.com/minimaxir/sdxl-experiments/blob/main/wrong_image_generator.ipynb).
+
 ## Notes
 
-- The intuitive way to think about how this LoRA works is that on training start, it indicates an undersirable area of the vast highdimensional latent space which the rest of the diffusion process will move away from. This may work more effectively than textual inversion but more testing needs to be done.
+- The intuitive way to think about how this LoRA works is that on training start, it indicates an undesirable area of the vast highdimensional latent space which the rest of the diffusion process will move away from. This may work more effectively than textual inversion but more testing needs to be done.
+- The description of this LoRA is very careful to not state that the output is objectively _better_ than not using LoRA, because everything is subjective and there are use cases where vibrant output is not desired. For most use cases, the output should be better desired however.
 - It's possible to use `not wrong` in the normal prompt itself but in testing it has not much effect.
 - You can use other negative prompts in conjunction with the `wrong` prompt but you may want to weight them appropriately.