flux klein 9b with prompt enchancer - 100% of understanding

this workflow, have prompt enchancer, so 100% of prompt understanding.
and i add a speed ups methods for increase up speed of generations.

just download the last version ever.

Prompt Enhancer for ComfyUI: clearer intent, more useful prompts

I created and adapted Prompt Enhancers for my FLUX.2 Klein 9B, Krea 2 Turbo, and Qwen Image 2.1 workflows. The goal is to take an idea written in natural language and turn it into a clearer, better-structured visual description for the image model.

Model used to enhance prompts

The enhancer uses Qwen3‑VL 8B in NVFP4 format through ComfyUI’s CLIPLoader and TextGenerate nodes. It receives the user's request along with model-specific instructions to identify the key elements of the scene—subjects, actions, positions, spatial relationships, setting, style, and visible text—and then write the final prompt in natural English prose.

Qwen3‑VL 8B is the auxiliary model that interprets and rewrites the request. It is separate from each image generator’s own encoder: the enhancer prepares the text, while the native encoder for FLUX, Krea 2, or Qwen Image 2.1 continues to handle image conditioning.

What the enhancer is designed to do

For short requests, it adds compatible visual details—such as environmental context, lighting, texture, or composition—to make the intended scene easier to render. For detailed requests, it is instructed to improve clarity without reworking the scene.

Explicit details should remain intact. If the request is “a dog on top of the table during a family dinner,” the final prompt should still include the dog, its position on the table, and the family dinner. Supporting details can enrich the description as long as they do not replace those key elements.

The enhancer is also instructed to respect the requested style, preserve visible text exactly as written, and avoid adding details that change the idea. The final prompt can be checked in the preview before image generation.

Workflow features

Each workflow includes a switch between the original and enhanced prompts (ORIGINAL / ENHANCED). This makes it easy to compare both paths or turn off the enhancer whenever you prefer.

All three workflows have instructions tailored to their respective image models. The way a scene is described should suit the generator being used, while the enhancer remains separate from each model’s native encoder.

I tested the results with the FLUX.2 Klein and Krea 2 workflows and was very happy with how well they followed the original idea. I also adapted the same approach for Qwen Image 2.1; that workflow still needs a generation test.

Made for ComfyUI. I’m sharing the approach and workflows for anyone who wants to try them and compare the results with their original prompts.