This is a simple setup I devised to help better understand the effects of the various conditioning combination methods. There are 2 fields for prompt: 1st and 2nd. The prompts will be fed into separate samplers to see how they look on their own. After this, there are 3 additional samplers that will combine the prompts in various ways: Concact, combine, and weighted average. There are previews for each type of conditioning combine, and a slider above the Average preview that allows the weight to be adjusted. I have renamed "Negative" to "Comparative Conditioning" since that is what it actually does. The node for negative prompt is labeled as "Leave this Blank" for good reason. Since we are trying to analyze the effects of conditioning combinations, it is best to limit the number of variables that can affect the system. My recommendation for this setup is to use prompts that are not related at all (IE, "1gir"l for 1st prompt, and "bowl of fruit" for 2nd prompt), so we can get a better idea of how the UNET is interpreting the embedding vectors. Uses AnythingEverywhere nodes and ComfyRoll. Use the ComfyUI Manager to download missing nodes.

Prompt Engeering Test Workflow: Combining Text Prompts

Fixed incorrect node names on image previews. Fixed global seed not applying same seed to all nodes. Added CFG control.

Files with SHA256: b4e7d731d031a759c6d4ef5e892b0bb078aad7a26ad92b89185ff798520a8494