This is my submission for the "Most Complex Workflow in the Universe" Challenge.π All-in-One Universal ComfyUI Workflow (Nodes 1.0 and 2.0 Ready)
A comprehensive, modular pipeline built for maximum flexibility and speed. Switch seamlessly between top-tier image architectures, edit modes, and intelligent prompt enhancement without rebuilding your graph.
π Workflow Overview & Major Updates
π¨ Supported Models & Pipelines
* Next-Gen & SOTA Models: Flux 3 Klein, Qwen Image 2.1, H3-Image, Ming Image 0.1, Ideogram 4, Z-Image, Krea-2
* Core Architectures: Flux, Flux 2, Flux 2 Klein, Qwen Image, Qwen Image Edit, Anima, SDXL / Pony
* Modes: txt2img, img2img, Inpaint, and Outpainting (available in Edit / Kontext mode)
* Format Support: Full compatibility with both .safetensors and .gguf checkpoints and text encoders
π§ Smart LLM Prompt Engine & Vision
* 1-Click Model Tier Switch: Toggle between a lightweight fast upsampler and a heavyweight 27B Vision/Text model without rebuilding the graph. Upsampler switch located in Value Presets block.
* Auto-Captioning & Enhancement: Built-in optical image scanner to extract detailed descriptions from reference images or expand short concepts into rich prompts.
* Text-to-Text Transformations: Turn off image processing in LLM group to instantly rewrite existing prompts (e.g., feed a sunny city prompt + instruction "Note to LLM: change to nighttime. Prompt: your daytime prompt").
* Vision-Guided Swaps (No Edit Models Needed!): Enable image processing in LLM group, load a reference (e.g., a man in a tuxedo), and instruct "replace man with an anthropomorphic tiger". The LLM extracts the scene and writes a pure T2I prompt for a tiger in the exact same outfit and pose.
* Real-Time Prompt Injection: Dynamically feed arbitrary instructions to the LLM (Note to LLM: ) directly from the user input fields while the i2i or t2i generation is active.
* Smart Auto-Bypass: The LLM block automatically sleeps to save time and resources if text input in LLM group is empty, no image is loaded, and the LLM toggle is disabled in your active i2i or t2i mode.
### β‘ Workflow & Navigation Features
* Nodes 2.0 Support: Clean, modern, and optimized node layout.
* Presets Selector: Instant configuration switching for different model families.
* Fast Keyboard Shortcuts: Jump across workflow groups instantly β try pressing 0, 1, 2, s, u, l.
* Flexible Reference Loading: Load single or multiple reference images for Edit and Krea-2 workflows.
### π οΈ Post-Processing, Control & Refinements
* Multi-LoRA Loader: Powered by Lora Manager for stacking and managing multiple LoRAs cleanly.
* ControlNet & Image Preprocessing (img2img): Prep your input image on the fly with built-in utility toggles crop face, change resolution, remove background). Then, activate a specific ControlNet preprocessor switch Pose, Depth, Canny, HED, Normal Map, or Lineart) to extract the map, and add a matching instruction to your prompt (e.g., "use input image as normal map reference").
* SD Upscale: High-res detailing pass (Recommended upscaler: [Acly/Omni-SR](https://huggingface.co/Acly/Omni-SR)).
* AfterDetailer & Face Swap: Automated face/detail restoration (Note: Face swap is currently experimental and still being refined).
π² JSON Bounding Box (BBox) Routing Guide: π²
1οΈβ£ π Split JSON: Route coordinates to the "Create Bounding Boxes (PPM)" node. Map fixed values (main object, style, etc.) to the "Build JSON Prompt (Ideogram)" node.
2οΈβ£ πͺ Auto-Parse: Click "Clear All" on the PPM node, then click "text to bbox nodes (no generation)". This parses JSON from user input field into both nodes without starting the actual image generation.
3οΈβ£ π€ Manual Elements: If the "Is manual elements" switch is ON, the visual boxes on the PPM node are ignored, and ONLY the text from the manual elements input field is used.
4οΈβ£ ποΈ Generation Controls: You can now run the generation by toggling "Bbox mode (ignore user prompt)" and/or "ignore bbox coordinates" as needed.
5οΈβ£ π§ Smart Appending: Even if "Bbox mode (ignore user prompt)" is DISABLED, your elements (from PPM or manual input) are STILL appended to the prompt. For non-JSON models, coordinates are automatically translated into natural language (e.g., "top right").
**v13.0 Update: The Next-Gen Core**
Current v13 status: Done. Beta testing in progress.
π *Complete Under-the-Hood Overhaul:** Familiar interface, entirely new engine.
β‘ *On-the-Fly Prompt Enhancement:** Seamlessly upgrade both T2I and I2I prompts in real-time.
π― *Supreme BBox Handling:** Fully optimized for Ming, Flux 3 Klein, and Ideogram.
π§ *New Model Integration:** Added native support for Prina AI and Minimax H3-Image models.
π *Flexible LLM Routing:** Choose between a dedicated node for direct LLM communication (GGUF format only) OR standard upsampling nodes for H3-Image, Flux 3, and Ideogram 4 (supports any format). You can also experiment with any text encoder or LLM model for your upsampling needs.
**v12.0 Update: Bbox attempt (failed)**
π― *Some BBox Handling:** Ming, Flux 3 Klein, and Ideogram.
**v11.0 Update: Universal Loader & Presets**
ποΈ *One Loader to Rule Them All:** Separate model groups are gone. Everything is now handled by a single Universal Loader and Preset Selector.
πΎ *Custom Presets:** Mix and match models, CLIPs, CFG, and steps. Just set your .safetensors or .gguf paths and hit "Save" or "Save As" to create your own recipe with a new name and settings.
π« *SDXL Override:** If the SDXL switch is ON, custom UNet, CLIP, and VAE settings are automatically ignored.
π¦ Use Krea2 edit only with this Lora
---
π§© Required Custom Nodes (which the built-in manager might not find)
Please ensure these are installed via ComfyUI Manager or manually cloned to avoid broken nodes:
π¦ [ComfyUI-Lora-Manager]- Handles LoRA recipes. (There is a "Save to recipe" action in the right-click menu of the LoRA list in the frontend).
π¬ [ComfyUI-LLM-text-processor] - Core LLM integration.
π² [ComfyUI-PPM] - Visual bounding box generation (Runs in 1.0 node mode).
π₯ [ComfyUI-OBVPM] - Value Presets, bundles and some unique bypass nodes.
---
πΎ Recommended LLM Models (By VRAM)
Note: Use the exact system prompts indicated for your hardware tier. Modern LLM-system-uni.txt can be used for any task.
**8 GB VRAM**
* [Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-Q8_0.gguf](https://huggingface.co/HauhauCS/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive/resolve/main/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-Q8_0.gguf)
* [mmproj-Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-BF16.gguf](https://huggingface.co/HauhauCS/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive/resolve/main/mmproj-Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-BF16.gguf)
**12+ GB VRAM**
* [Gemma-4-E4B-Uncensored-HauhauCS-Aggressive-Q6_K_P.gguf](https://huggingface.co/HauhauCS/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive/resolve/main/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive-Q6_K_P.gguf)
* [mmproj-Gemma-4-E4B-Uncensored-HauhauCS-Aggressive-f16.gguf](https://huggingface.co/HauhauCS/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive/resolve/main/mmproj-Gemma-4-E4B-Uncensored-HauhauCS-Aggressive-f16.gguf)
* (Also compatible with Qwen3.6-35B-A3B versions)
**16 GB VRAM**
* [Qwen3.8-27B-Uncensored-IQ4_XS.gguf](https://huggingface.co/JonathanColetti/Qwen3.8-27B-Uncensored-GGUF/blob/main/Qwen3.8-27B-Uncensored-IQ4_XS.gguf)
* [mmproj-Qwen3.8-27B-Uncensored-F16.gguf](https://huggingface.co/JonathanColetti/Qwen3.8-27B-Uncensored-GGUF/blob/main/mmproj-Qwen3.8-27B-Uncensored-F16.gguf)
**24 GB VRAM**
* [Qwen3.8-27B-Uncensored-IQ5_K_M.gguf](https://huggingface.co/JonathanColetti/Qwen3.8-27B-Uncensored-GGUF/blob/main/Qwen3.8-27B-Uncensored-Q5_K_M.gguf)
* [mmproj-Qwen3.8-27B-Uncensored-F16.gguf](https://huggingface.co/JonathanColetti/Qwen3.8-27B-Uncensored-GGUF/blob/main/mmproj-Qwen3.8-27B-Uncensored-F16.gguf)
---
## π Folder Structure
Place your models and system text files exactly like this:
π ComfyUI/
βββ π models/
β βββ π LLM/
β βββ Qwen3.8-27B-Uncensored-IQ4_XS.gguf
β βββ mmproj-Qwen3.8-27B-Uncensored-F16.gguf
β βββ π prompts/
β βββ βββ LLM-system-uni.txt
β βββ βββ LLM-system-Old.txt
β βββ βββ LLM-system-Old-sdxl.txt
β βββ βββ Minimax-H3-image-i2i.txt
β βββ βββ Minimax-H3-image-t2i.txt
βοΈ Smart Prompting Guide (v13.0 Universal Logic)
**Prompt Anatomy:** User prompts are now divided into two parts:
1. **Fixed Prompt:** Stays exactly as written at the very beginning (NEVER processed by the LLM).
2. **User Prompt:** Handed off to the LLM for enhancement, styling, or formatting (if the LLM is enabled).
**π₯ Minimax H3-Image Templates i2i.txt & t2i.txt):**
Included in the prompts folder are specialized, compact system prompt templates explicitly designed for the Minimax text encoder upsampler.
* Minimax-H3-image-t2i.txt: Optimized for generating scenes from scratch.
* Minimax-H3-image-i2i.txt: Tuned for reference-conditioned image-to-image transformations.
(Both templates are built to support reasoning models and work perfectly in tandem with the workflow's custom Regex nodes, which automatically strip <think> tags from the final output before sending it to the generator).
---
ποΈ How to Command the AI (The 4 Routing Modes) with LLM-system-uni.txt
The universal system prompt acts as a strict translation engine. Do not over-explain or ask the AI to "keep the background the same" β just state the final result you want to see. The AI will automatically lock into one of 4 modes based on your input:
**1οΈβ£ MODE A: Default Generation (Full Prose)**
* **Trigger:** Just type your idea (with or without an uploaded image).
* **Output:** The LLM auto-generates a rich 200β400 word paragraph defining lighting, camera angles, kinematic poses, and textures, keeping everything in sharp focus. Automatically translates simple styles (e.g., "cinematic photo", "Pixar 3D") into complex anatomical and optical vocabulary.
**2οΈβ£ MODE B: Visual Delta Edit [edit])**
* **Trigger:** Upload an image and type [edit] followed by your command (e.g., [edit] make her hair blonde).
* **Output:** The LLM outputs a concise 15β60 word imperative instruction (InstructPix2Pix style). It explicitly details ONLY the modified elements and their lighting integration, without hallucinating full-scene descriptions.
**3οΈβ£ MODE C: Spatial & JSON Parsing [bbox])**
* **Trigger:** Include [bbox] in your prompt OR feed an existing JSON {...} structure into the input.
* **Output:** PURE JSON mode. The AI generates or modifies a comprehensive JSON scene layout (compatible with Ideogram, Flux 3 Klein, or PPM nodes). It respects existing "bbox" coordinates while intelligently rewriting the "desc" and high-level descriptions for each element.
* (Pro tip: Combine an existing JSON with [edit] instructions to spatially edit specific regions!)
**4οΈβ£ MODE D: Danbooru/SDXL Tags [sdxl])**
* **Trigger:** Include [sdxl] anywhere in your prompt.
* **Output:** A single-line, comma-separated list of 70β150 tags (perfect for Pony, Illustrious, or traditional SDXL models) strictly ordered by quality, medium, subject, and background.
**π Forcing Canvas Size (Applies to Mode C & D):**
Add TARGET CANVAS: W:H anywhere in your text (e.g., TARGET CANVAS: 16:9 or 3:4) to force specific aspect ratios into the JSON or tag outputs.
Description
Better LLMs, more models, better bboxes







