Overview

A complete, working ComfyUI workflow that combines local vision-language captioning (via Ollama + Qwen3-VL 8B) with a high-quality text-to-image generation pipeline.

Feed it any image and it will automatically write a detailed, natural-language recreation prompt, then generate a new image from that prompt. You can also switch to fully manual prompting at any time. Includes upscaling, memory cleanup, and organized saving.

This is the exact workflow I’ve been using for character-consistent recreation and prompt extraction (especially useful for the Krea-style detailed spatial/pose prompting).

Key Features

• One-click Image → Detailed Prompt Uses OllamaGenerateV2 + Qwen3-VL:8B with a carefully tuned system prompt that forces clean, accurate, natural-language descriptions (no tags, no thinking process in the output).

• Manual / Auto Switch TwoWaySwitch node lets you instantly choose between the Ollama-generated prompt or your own manual prompt.

• Full Generation Pipeline Ready-to-run text-to-image path with ModelSamplingAuraFlow, proper VAE, conditioning, and latent handling.

• Upscale Path Separate upscale group with SeedVR2-style restoration + saving.

• Memory Management Built-in VRAMCleanup, RAMCleanup, and easy cleanGpuUsed nodes so the workflow stays stable during long sessions.

• Organized Saving LoraManager Save Image nodes with date-based folders (%date:yyyy-MM-dd%\CFH and Upscale variants).

Requirements

Must have:

• Ollama running locally

• Model: qwen3-vl:8b (pull it with ollama pull qwen3-vl:8b)

• ComfyUI with the following custom node packs:

  • comfyui-ollama

  • controlaltai-nodes

  • comfyui-kjnodes

  • comfyui-easy-use

  • comfyui_memory_cleanup

  • comfyui-lora-manager

  • comfyui_tinyterranodes

  • RES4LYF (for TextConcatenate)

  • Any nodes required by your upscale/SeedVR2 path

How to Use

1. Load the workflow.

2. Make sure Ollama is running and qwen3-vl:8b is available.

3. Drop an image into the Load Image node that feeds the OllamaGenerateV2 node.

4. Set the TwoWaySwitch:

   • Position 1 = use the auto Ollama prompt

   • Position 2 = use your manual prompt

5. Queue the prompt.

6. The workflow will generate the caption → create the image → optionally upscale and save.

Notes

• The Ollama system prompt is already tuned for maximum accuracy and clean output (no chain-of-thought leakage).

• Everything is grouped and color-coded for easy navigation.

• Designed for iterative character work and precise pose recreation.

Feel free to use, modify, and share. If you improve the Ollama system prompt or add better control nets, drop a comment — always happy to see upgrades.

Enjoy.