In this workflow, I’ve built an intelligent ComfyUI setup that automatically improves user prompts to better suit the Flux1 Kontext-Dev editing system — a cutting-edge tool for image-to-image editing .
📘 Reference: Flux1 Kontext-Dev Official Guide
🎯 Goal
Flux1 Kontext-Dev relies heavily on clear, rich, and well-structured prompts to guide the editing process. However, many users provide short or vague prompts, leading to poor results.
This workflow solves that by integrating a local large language model (LLM) using Ollama, which rewrites simple prompts into descriptive, detailed prompts tailored for effective image editing.
⚙️ How the Workflow Works
User Inputs:
An image for editing.
A simple or vague text prompt describing the desired change.
Ollama Integration (LLM for Prompt Enhancement):
The prompt is passed to Gemma-3, a vision-enabled LLM running locally via Ollama.
The model rewrites the prompt into a more expressive and visually descriptive version.
Enhanced Prompt → Flux1:
The improved prompt is fed into the Flux1 Kontext-Dev nodes along with the input image.
Flux1 then performs context-aware image editing based on this high-quality prompt.
📦 Requirements
To run this workflow, you need the following components:
✅ 1. Ollama
A powerful local runtime for LLMs and vision models.
🔗 Download and install Ollama:
https://ollama.com/download
✅ 2. Vision Model: gemma3
Use a multimodal (vision + language) version of Gemma 3 depending on your system’s VRAM:
👉 Model Page:
https://ollama.com/library/gemma3
ollama run gemma3🔥Uncensored Model:
https://ollama.com/huihui_ai/gemma3-abliterated
ollama run huihui_ai/gemma3-abliterated⚠️ Make sure you're using the multimodal (vision) variant of Gemma 3 to ensure it can process image-based prompts in ComfyUI.
✅ Key Benefits
Improved editing accuracy from even simple input prompts.
Local-first, privacy-safe setup using Ollama and ComfyUI.
Flexible model choices depending on your hardware.
💡 Example
Input prompt:
"change the style to realistic"
Enhanced prompt via Gemma-3:
"Change the image to a photorealistic rendering, with accurate lighting, textures, and details, while preserving the subject’s facial features, pose, and the existing composition."
🌍 Multilingual Prompt Support
This workflow supports prompts in any language, including Arabic, and automatically translates them into expressive English prompts that Flux1 can interpret.
💬 Example:
Input (Arabic):
"حول الستايل إلى حقيقي"
Enhanced Output (English):
"Change the image to a photorealistic rendering, with accurate lighting, textures, and details, while preserving the subject’s facial features, pose, and the existing composition."
This makes the workflow highly accessible to non-English speakers while still benefiting from professional-grade prompt enhancement.
🧩 Workflow Versions
There are two versions of this workflow available:
🔹 Basic Version
Designed for ease of use.
Supports 1–2 input images.
🔸 Advanced Version
Supports up to 4 input images.
Includes upscaling at the end of the pipeline.
Built for professional-quality outputs.
Based on a modified version of this original workflow from Civitai:
👉 https://civarchive.com/models/618578?modelVersionId=1956938