Workflow-Thinking-2.0
🚀 [Update] workflow-Thinking Integrates Gemma 4 12B IT QAT Thinking Mode!
💡 What’s New:
Enhanced Output Quality: By leveraging this new thinking mode, the model now delivers significantly better reasoning capability, deeper insights, and higher-quality responses for complex tasks.
Longer Inference Time: Please note that due to the deeper reasoning and "thinking" process, the model's inference/generation speed will be slightly slower than before.
⚖️ Our Take: This is a classic "quality over speed" trade-off. If your tasks require high precision and rigorous logic, we highly recommend giving this new workflow a try! We will continue to optimize inference efficiency to better balance speed and performance.
Give it a spin and let us know your thoughts in the feedback channel!
⚠️Please read the v3.7.1 update announcement to download the corresponding nodes and models.
Workflow Update: anima-base-v3.7.1
The updated workflow (anima-base-v3.7.1) has been officially re-uploaded! This release builds upon the major overhauls from v3.0 and v2.5, bringing refined prompt templates, new shading styles, and critical backend stability.
🌟 What's New in v3.7.1
New Style Prompt Templates: Added dedicated prompt templates for Cel-shading, Soft Shading (Thin Painting), Semi-Painterly, and Painterly (Thick Painting).
Universal Prompt Interrogation: Added a general prompt reversal/interrogation template to streamline your tag extraction.
Prompt Fixes: Sneaked in various optimization fixes to existing prompt templates for smoother generations.
📦 Recommended Model Upgrade
To get the best results with the new interrogation templates, it is highly recommended to download the latest LLM prompt interrogation model:
👉 SC117/gemma-4-12B-it-heretic-QAT-GGUF · Hugging Face
🛠️ Complete Installation & Setup Guide
If you are upgrading from an older version (pre-v3.0), you will need to install two core custom node ecosystems and place your LLM model in the correct directory.
1. Install Required Custom Nodes
Open your terminal, navigate to your ComfyUI custom nodes directory, and run the following commands:
Bash
cd ComfyUI/custom_nodes
# Node 1: LLM Text Processor (Required for Gemma 4)
git clone https://github.com/KingManiya/ComfyUI-LLM-text-processor.git
# Node 2: Anima Style Explorer (Required for Style Palettes & Switcher)
git clone https://github.com/mizukir0418-gif/ComfyUI-Anima-Style-Explorer-Prompt.git
(Alternatively, you can download the ZIP files directly from GitHub and extract them into your custom_nodes folder, ensuring no -main suffixes remain in the folder names).
2. Place the LLM Model
Model Link: Download
gemma-4-12b-it-uncensored-GGUF(or the recommendedheretic-QATvariant) from Hugging Face.Storage Path: Place the
.gguffile strictly inside:ComfyUI/models/LLM/
⚠️ Critical Troubleshooting: Fixing First-Time Loading Errors
🚨 IMPORTANT: Because Gemma 4 utilizes a brand-new architecture, the older default backend built into the LLM node will throw an error (e.g., unknown projector type or inference failed) on its first run.
If your ComfyUI crashes or errors out, follow these manual steps to update the backend:
Download the Latest Backend: Go to the official llama.cpp GitHub Releases.
Select the Right File: Download the latest release zip corresponding to your GPU setup (e.g.,
Windows x64 (CUDA 13) - CUDA 13.3 DLLsor the CUDA 12 variant).Locate the Target Folder: Navigate here inside your ComfyUI directory:
ComfyUI/custom_nodes/ComfyUI-LLM-text-processor/vendor/llama.cpp/b8840/win-x64-cuda13/Replace the Files: Extract the downloaded zip and copy-paste all files directly into the
win-x64-cuda13folder, overwriting the existing ones.Restart: Restart your ComfyUI backend.
🚀 Core Features Recap (v2.5 – v3.0 Cumulative)
If you missed the previous major updates, here is what is now completely integrated into your v3.7.1 pipeline:
🧠 Gemma 4 Integration & Advanced Tag Reversal
Handles highly detailed, Danbooru-style descriptive captioning, advanced anatomical logic conversion (perfect for gender-bending and complex character transformations), and precise prompt refinement. (Includes optimized tag-interrogator prompt templates for terms like 'femboy').
🔥 Core Node: [Anima 2B Dynamic Group Switcher]
Say goodbye to manually toggling dozens of rgthree groups!
Automatic Group Name Detection: Connecting a slot to a style group automatically reads the group's title and re-labels the input port natively.
Automatic Bypass Matrix Control: Selecting an option from the
active_styledropdown automatically activates that group while instantly forcing all unselected groups into Bypass mode (turning them gray).Zero-Crash Backend: Built-in validator completely eliminates native ComfyUI
Value not in listruntime errors when switching styles mid-generation.
🎨 The 3 Masterpiece Palette Templates
Integrated JoyCaption-optimized aesthetic templates to instantly elevate your textures:
Palette TemplateVisual FormulaLighting & Texture❄️ Muted / MorandiLow-saturation tones (dusty rose, sage green, muted grayish blue) over soft grey undertones. Strips away harsh digital neon colors.Diffused overcast lighting, minimal contrast, zero harsh shadows. Gallery-grade aesthetic.🍂 Retro Vintage / 90s AnimeForces a 1990s cel-shaded aesthetic with rich, warm sepia undertones and faded nostalgic colors.Injects optical vintage film grain and subtle chromatic aberration (lens edge color-bleeding).🖤 Color Split / Selective ColorLocks environment, background, skin, and primary attire into a rich, high-contrast monochrome grayscale.Isolates a single visual focus item (e.g., red lips, glowing jewelry) in high-saturation color.
🧩 The Expanded 4-Node Ecosystem
Your workspace now operates as an industrial-grade production pipeline:
Anima 2B Dynamic Group Switcher: Automated workflow traffic controller.
Artist Pre-Config Outline: Handles macro-category wrapping and multi-artist nested weight blending.
Anima 2B Prompt Template 1: Core parallel multi-weight artist channel generator.
Anima 2B Prompt Template 2: High-end modifiers, time eras, structural evaluation, and prompt quality scoring.
🚨 Final Upgrade Notice
Because this ecosystem relies heavily on an integrated Frontend JavaScript architecture, you MUST hard-refresh your browser (Press Ctrl + F5) after updating and restarting ComfyUI. Failing to clear your browser cache will cause interactive dropdown menus and buttons to malfunction.
Enjoy the new version, and feel free to share your generations or drop any questions in the comments below!
Description
## 🌟 Shoutout to Google: Why Gemma 4 12B is an Absolute Game-Changer for Prompt Reversal!
I have to take a moment to give massive credit to Google for open-sourcing the Gemma 4 12B model. After extensive testing and integrating it into this v3.0 workflow, I am genuinely blown away by how incredibly powerful and smart this compact model is.
If you are using raw WD14 tags and trying to reverse-engineer them into rich, natural language captions for models like Anima-base, Gemma 4 12B is the ultimate tool. Here is a breakdown of why it completely destroys older models when executing complex prompt templates:
---
### 🧠 Why Gemma 4 12B Smashes Prompt Reversal & Captioning:
* Flawless Logic & Strict Rule Adherence: Most 7B to 12B models easily get confused when you feed them heavy, rule-based templates (like our Identity Anchoring and Anatomy Locks). They suffer from token bleeding and mix up characters. Gemma 4 behaves like a 70B+ model. It reads the strict constraints, memorizes who is who, and executes the output without a single logical slip-up.
* Insane Dynamic Tag Translation: When we feed it a chaotic list of raw Danbooru tags, Gemma 4 doesn't just copy-paste them. It acts as an intelligent translator. For gender-bending/crossdressing workflows, it flawlessly converts complex, highly specific female-centric tags into correct, biologically consistent male/femboy descriptions (e.g., smoothly re-mapping interactions like tribadism into frottage, or handling active footjobs/handjobs with precise physical motion instead of turning them into passive states).
* Zero Creative Filters (Uncensored GGUF): Thanks to the uncensored community fine-tunes, Gemma 4 completely sheds its corporate restrictions. It allows for absolute creative freedom, capturing hyper-detailed, explicit visual states, skin textures, fluid transparencies, and complex anatomical angles with zero hesitation or censorship refusal.
* Perfect Danbooru-to-Prose Synergy: It understands the exact nuance behind Danbooru tags. It can effortlessly read tags like large_breasts, erection_under_clothes, or veiny_penis and weave them into a single, seamless, continuous natural language paragraph. It heavily emphasizes micro-details—like the precise way fabric stretches over an erection—making the final caption incredibly potent for AI image generators.