v2 Update - Updated Custom Nodes, optimised file path utilisation, added Detail Daemon Sampler which surprisingly adds lots of details at a very low cost before sampling, added FL Upscale for faster upscale time - but less high quality than Ultimate SD Upscale.

I've spent many days refining this workflow so that you don't need to manipulate many things. I've made it as understandable as possible by explaining every node in the workflow with Notes next to them. Full explanation of every group is below.

Please post an image of your result as I receive Buzz⚡I'm curious to see your results!
Any tip is appreciated, either by Buzz⚡ or coffee☕.

?Description

This workflow will randomly select a reference image from a folder, use Florence2 to describe it, and use the description as part of the prompt. To make the result more interesting, this workflow mixes a detail Lora + 2 random style Loras. Also, wildcards will choose 1 out of 100 fantasy backgrounds, 94 camera angles, and 39 camera lenses on every queue.

?How this Workflow Works

1. ? Randomize Lora

2. ?Style + Detail Enhance Lora

3. ?️Text Styles Alternative

4. ? Wildcards - Background, Camera angle & lens

5. ?️Image Reference Loader

6. ✍️FLUX Prompt

7. ?Upscale

1. ?Randomize Lora

• Enter the path to your Flux Lora folder in the Lora Folder Directory.

• If you've modified the default ComfyUI "loras" name, update the CR Split String value below.

• If you don't have any preview images of your models, you can disable/bypass the blue nodes.

2. ?Style + Detail Enhance Lora

• In the Style Lora, in the lora_name field, you will need to manually choose one of your lora. Otherwise, you might get an error as it won't find the current Lora

• In the Detail Enhance Lora, you will need to choose a detailer Lora. One that adds more detail. The FaeTastic Detailer is a great one.

• I've left the option to have 2 detailers Lora as sometimes some detailers add more details around in the landscape and some that enhance the subject detail.

• If you don't want to use the Randomize and want fixed Lora styles, you can bypass the Randomize. If you have preview images, you can toggle the enable_preview to display them.

3. ?️Text Styles Alternative

• If you don't have many Style Loras, this node is a great alternative.

• You can select many Styles that will output a chain of text prompts to describe the style.

4. ?Wildcards - Background, Camera angle & Lens

• These wildcards will randomly choose 1 out of 100 fantasy backgrounds, 94 camera angles, and 39 camera lenses on every queue.

• You can modify the txt files or create your own and even duplicate the Chibi-Nodes for more randomness. Just don't forget to connect the text output to a text concatenate.

• I've provided the txt files that you will need to store in the chibi-wildcards custom nodes directory :

ComfyUI\custom_nodes\ComfyUI-Chibi-Nodes\extras\chibi-wildcards

5. ?️Image Reference Loader

• This is where the majority of the magic comes from. One of your images from your folder will be loaded and described by a Large Language Model (Florence2). This description will then be used in the prompt.

• Load Image node is where you enter the directory of your image reference folder. I have 270 images in my folder located in ComfyUI\input\Ref.
  ⚠️For this to work, must add the "\" at the end of the path⚠️

• If you find the description is too long, in Florence2Run node, you can modify the task to "detailed_caption".

• DownloadAndLoadFlorence2Model node will automatically download the Florence-2-Flux-Large model in the LLM model folder and you run this workflow for the first time. When you have the model, it won't download again. It will load it instead.

ComfyUI\models\LLM\Florence-2-Flux-Large

6. ✍️FLUX Prompt

• All information is gathered here. You can see the total prompt on the right.

• Wildcard Processor node is the only node that you can manually edit the prompt. You can precise your own theme in here.

• Seed Everywhere node is the seed used for all nodes that require randomness.

• If you want to re-generate the same prompts and image reference, use the seed below and replace it with the field.

• Unfortunately, the only randomness that isn't consistent is the Randomize Lora. I'm waiting for the DirGir Node to fix the random issue in the index. Without my current workaround, you would need to manually modify the index number every time you load the workflow and would need to re-enter the folder directory if you've duplicated it.

7. ?Upscale

• This upscales the image by 2x.

• You can disable the Upscale group to generate multiple images faster by using the Mute.

• In the Load Upscale Model node, I use the Ultrasharp model, which can be found on Huggingface.

• If the Image Comparer (rgthree) isn't working, look at this issue fix.

• Place the model here:

ComfyUI\models\upscale_models

If you have any questions, feel free to ask in the Discussion section ?

描述:

More detail, optimised and up to date Comfy

训练词语: