This is for creating images in FLUX and detailing them with Z_Image_Turbo. You can also compare the outputs in this workflow. While this takes more time than running either model independently, the final output is a much more authentic looking image.

Image Comparison:

After Detailing in Z_Image_Turbo:

Before Detailing in FLUX:

Prompt (for both models): A strong, fluffy grey striped short haired cat presses itself into the fluffy carpet preparing to pounce on a toy ball. The cat's eyes are fully dialated, it's sole focus the ball.

Note that the prompt is included in the workflow.

How to use

1. Input a prompt into the PROMPT Box in the constants section. I recommend tuning the prompt for Z_Image_Turbo over FLUX as FLUX works pretty well with Z_Image_Turbo tuned prompts but not vice versa.

2. Select your desired FLUX and Z_Image_Turbo based models

3. Configure the Width and Height of your generation. Values past 2048 could have issues

4. Run the generation and compare between raw FLUX vs with Z_Image_Turbo Providing detailing

Suggested Modifications

Workflow Modifications

1. Add LoRAs to FLUX/Z_Image_Turbo

2. Add a Save Image node to save the generation automatically at the end of the workflow

Dials to tune

1. Steps on Flux can be increased or lowered depending on initial image quality, a little blur is generally fine if you are adding a lot of detail

2. denoise on the Z_Image_Turbo KSampler can be tuned as desired. Lower values incorporate more of the image from FLUX while higher values add more detail from Z_Image_Turbo

3. Tune individual prompts for FLUX and Z_Image_Turbo

What it does

1. PROMPT provides a prompt to both FLUX and Z_Image_Turbo. While FLUX and Z_Image_Turbo have different preferred formats, I found that prompts tuned for Z_Image_Turbo performed well enough in FLUX. It will show you the prompt in the Chosen Prompt Node.

2. FLUX Generates a base image. FLUX is good at image diversity from a single prompt allowing you to get varied images with different seeds for the same prompt (e.g. different people in portraits, different landscapes) Z_Image_Turbo tends to disregard the seed for the most part and just create one image per prompt, especially when the prompt is detailed. FLUX therefore provides diverse and interesting base images, but with less detail than Z_Image_Turbo can provide.

3. Z_Image_Turbo adds detail. The amount of detail added is approximately inversely proportional to the amount of FLUX image that is retained in the final output (at least between 0.00-0.25 denoise). Denoise values between 0.10-0.20 appear to provide the best results.

4. The images are presented for comparison allowing you to AB test them or manually inspect the detail added by Z_Image_Turbo.

Used Plugins and what they do

All plugin nodes are optional and can be removed; however, they do provide quality of life improvements.

1. rgthree-comfy

   1. Power Prompt (rgthree) allows specifying a single prompt including random selection values (e.g. The {tiger|dog} chases the sheep) to populate the prompts of both FLUX and Z_Image_Turbo easily. If you remove this node, just make sure to include prompts for both FLUX and Z_Image_Turbo.

   2. Image Comparer (rgthree) allows easy comparison of the standalone FLUX image vs the image after Z_Image_Turbo detailing. You can remove this, just make sure to add image preview nodes to see the images side by side.

2. comfyui-custom-scripts

   1. Show Text 🐍 allows seeing the chosen prompt with all random values determined. This allows you to check for prompt adhesion for a particular run, especially if you have many random values.