Flux Tweak Clip Detailers版本v1.1 (ID: 1184731)

Flux with CLIP Attention and Detailers

This is my current workflow for Flux, designed to improve prompt adherence and text quality using CLIPAttentionMultiply adjustments. By tweaking these values, you can enhance image coherence to match your prompt more accurately. While it often increases image quality, experimentation is key—it can sometimes yield unexpected results. Play around and have fun!

What’s happening under the hood?

• QKV Explained (Query - Key - Value):

  • Q (Query): Determines how strongly tokens influence each other within a sentence.

  • K (Key): Assigns weights to tokens from the input text (words or sub-words).

  • V (Value): Controls the intensity of attention applied to input tokens.

Additional Enhancements

• I’ve included detailers for various body parts to refine and enhance specific areas of the generated images, ensuring a more polished and detailed output.

Feel free to explore, tweak, and experiment with the settings to see what works best for you!

Custom Nodes

• Anything Everywhere? /https://github.com/chrisgoringe/cg-use-everywhere

• Bookmark (rgthree) /https://github.com/rgthree/rgthree-comfy

• FaceDetailer /https://github.com/ltdrdata/ComfyUI-Impact-Pack

• Fast Groups Bypasser (rgthree) /https://github.com/rgthree/rgthree-comfy

• Image Comparer (rgthree) /https://github.com/rgthree/rgthree-comfy

• InjectLatentNoise+ /https://github.com/cubiq/ComfyUI_essentials

• JWInteger /https://github.com/jamesWalker55/comfyui-various

• Power Lora Loader (rgthree) /https://github.com/rgthree/rgthree-comfy

• ProjectFilePathNode /https://github.com/MushroomFleet/DJZ-Nodes

• SAMLoader /https://github.com/ltdrdata/ComfyUI-Impact-Pack

• SaveImageWithMetaData /https://github.com/edelvarden/ComfyUI-ImageMetadataExtension

• Seed Everywhere /https://github.com/chrisgoringe/cg-use-everywhere

• UltimateSDUpscale /https://github.com/ssitu/ComfyUI_UltimateSDUpscale

• UltralyticsDetectorProvider /https://github.com/ltdrdata/ComfyUI-Impact-Subpack

• UnetLoaderGGUF /https://github.com/city96/ComfyUI-GGUF

bbox

• face, hand, person, fashion: https://huggingface.co/Bingsu/adetailer/tree/main

  • Folder: models/ultralytics/bbox/or models/adetailer

• Eyes: https://civitai.com/models/178518/eyeful-or-robust-eye-detection-for-adetailer-comfyui

  • Folder: models/ultralytics/bbox/or models/adetailer

• Breasts: https://civitai.com/models/138918/adetailer-after-detailer-female-breast-model

  • Folder: models/ultralytics/bbox/or models/adetailer

• Private bits: https://huggingface.co/AunyMoons/loras-pack/tree/main

• Folder: models/ultralytics/bbox/or models/adetailer

Understanding YOLO Models and Which One to Pick

File Naming Convention

• Version Number: The number in the file name indicates the version of the YOLO model (e.g., YOLOv5, YOLOv8).

• File Type: The ".pt" extension signifies a PyTorch file, which contains the trained model ready for use.

• Model Variant: The version number is often followed by a letter, typically "s" or "n," denoting the model variant.

Model Variants Explained

1. Small ("s") Variant:

   • Optimized for a balance between speed and accuracy.

   • Compact model that performs well without being as resource-intensive as larger versions.

   • Suitable for environments with moderate computational resources.

2. Nano ("n") Variant:

   • Designed for very limited computational environments.

   • Prioritizes speed and efficiency, making it faster than the small variant.

   • Sacrifices some accuracy to achieve faster performance.

Choosing the Right Model

• "s" (Small): Ideal for scenarios where you need a good trade-off between speed and accuracy. Use this version if you have moderate computational resources and can tolerate a slightly larger model size.

• "n" (Nano): Best for resource-constrained environments where speed and efficiency are critical. Choose this version if you prioritize faster inference times and can accept reduced accuracy.

Both the small and nano variants are scaled-down versions of the original YOLO model, tailored to different levels of computational resource availability. Select the variant based on your specific use case and hardware limitations.

~~ Kiko!

描述:

Update:

• Fast Groups Bypasser instead of Muter, so the image passes through to the upscaler. As requested by user: @ianmichaelsty250

• documentation / links for nodes and models used.

训练词语: