This workflow converts any non-realistic images to realistic using Pony XL.
🆕 What's New in V3
Welcome to Pony-Anime-Cosplay-v3 workflow! Below are the key improvements and updates designed to optimize the whole workflow, improve facial recognition, and more features.
📈 Refiner
Added SDXL Refiner for more realism regardless of mode - uses JuggernautXL
Utilizes InsightFace and FaceID PLUS V2 for facial detection and face swap.
Uses Tile or Canny CNet for more accuracy - switch between two depending on the refiner output - defaults to Tile.
Switch between initial generated face (from Pony) or an image input.
⚠️ Note: Face detection may occasionally fail or detect the wrong face. If results are inconsistent, try different inputs (main and/or face image).
🌾 Film Grain Integration
Uses VRGameDevGirl Fast Film Grain (set to 0.02) to add more realistic output.
⚙️ Configuration Changes
💡 These are based on personal preferences. Feel free to adjust them to fit your workflow needs.
⏭️ Moved Highres-Fix to Refiner - it doesn't make sense to put on the first generation.
❌ Disabled cinematic photography... Lora.
➕ Added separate seed node to each detailer for more flexibility. I use it for testing purposes but can be useful if you want to regenerate a certain detailer.
🔄 Changed values for each CNet nodes.
➕ Added Additional Promp which appends to the end of final prompt. This is already available in the text concat, but due to the node placements, i just added it as a connection to the 6th concat. This allows us to be more precise, for example, character names if it wasn't added, or other charactre features missed by the tagger.
🧰 Quality-of-Life Improvements
🧩 General node rearrangement for a cleaner layout.
🔗 Updated Modes & Download Links – including updated default settings.
➕ Added new modes!
📖 Always start with the USAGE info inside the workflow.
Description
🆕 What's New in V2
Welcome to Pony-Anime-Cosplay-v2 workflow! Below are the key improvements and updates designed to enhance facial recognition, image detail, and overall usability.
🎭 IP Adapter FaceID (V2)
Utilizes InsightFace and FaceID Portrait for facial detection and embedding.
Enhanced facial recognition for features and expressions (FaceID Mode).
⚠️ Note: Face detection may occasionally fail or detect the wrong face. If results are inconsistent, try different inputs (main and/or face image).
🔄 Face Swap Mode
Enables input of a face image to perform face swapping operations.
⚠️ Note: Accuracy depends on the quality of the source image and facial alignment.
✋ Hand Detailer
Uses the same detailing method as the Face Detailer, now extended to enhance hand quality.
⚙️ Configuration Changes
💡 These are based on personal preferences. Feel free to adjust them to fit your workflow needs.
❌ Disabled Point and Shoot LoRA.
➕ Added
CyberRealistic_Negative_PONY_V2-negembedding to the negative prompt.🔄 Changed
combine_embedsin IPAdapter fromconcat➡️norm averagefor a more balanced blend of style and composition.📐 Adjusted image resize from 1024px ➡️ 1152px based on CyberRealistic Pony training data image's long edge.
🖼️ Set ControlNet image sources to Image Resize node (eliminates need for separate resize nodes in ControlNet).
✅ Enabled ControlNet in the Highres-Fix node for improved accuracy during upscaling - uses Canny.
🧰 Quality-of-Life Improvements
❌ Removed OpenPose, which was often unreliable in pose detection.
🔁 Replaced CR ControlNet Stack with OOB ControlNet Loader – simplifies UI.
🔗 Decoupled Canny and Depth ControlNets for more flexible use.
🧠 Reordered Detailer and Upscaler – eye detailer now runs before upscaling.
👁️ Removed Eye Detailer from the group bypasser.
🗂️ Removed Batch Node – never used.
🧩 General node rearrangement for a cleaner, more compact layout.
🔗 Updated Modes & Download Links – including updated default settings.
📖 Always start with the USAGE info inside the workflow.
▶️ Video Tutorial to follow.
FAQ
Comments (5)
lORA link pls
HighRes-Fix Scrip It's burning red. I can't install the NODE, what can I do?
Just replace it with a resize image and sampler node. hires-fix was a hack back in the day in a1111 and somehow it was ported over to comfyui even though we have way better tools available to upscale
It should be part of efficient nodes, but if you have an error message you can share, i might be able to help.
Auusming you have the same problem I did: I had to select Fix node (recreate) then change use_same_seed from true to false and the seed to 0. That got it working.