comfyui_prompt_bench

Prompt Bench

A small ComfyUI helper for iterative prompt improvement with Ollama.

It lets you simply:

• score an image, make notes and shape the result with structured ratings.
• turn feedback into a revised prompt.
• keep a chat-style history per session that is selectable using a seed.
• save prompt/rating snapshots to disk under comfyui/users/prompt_bench folder.

Nodes

PBTextModelChat

Generates the next prompt revision through Ollama chat.

Inputs:

• session_seed: session key; reuse it to continue the same conversation, change it to start a new one
• system_prompt: initial instruction for the model
• rating: structured feedback from PBRating
• model: Ollama model name
• text (optional): extra user input

Outputs:

• enhanced_prompt
• model
• text
• chat

PBRating

Creates a structured rating payload.

Inputs:

• notes
• prompt_adherence
• aesthetics
• technical_quality
• lighting
• detail_richness
• coherence

Output:

• rating

PBStore

Saves prompt runs as JSON files.

Inputs:

• prompt
• rating
• model
• text
• images (optional, currently unused)

Behavior:

• saves data under the ComfyUI user folder in prompt_bench/store/<text-based-id>/

PBListChatView

Formats chat history into readable text in the UI.

Inputs:

• chat
• text

Output:

• formatted chat transcript

Basic Workflow

1. Rate the current result with PBRating.
2. Send that rating into PBTextModelChat.
3. PBTextModelChat uses the current session history plus the new rating feedback to ask Ollama for a better prompt.
4. Send the generated prompt, rating, model, and text into PBStore to save the run.
5. Send the chat output into PBListChatView to inspect the session history.

Session Behavior

• Same session_seed = continue the same chat session
• Different session_seed = start a fresh session

Notes

• Chat memory is stored in memory while ComfyUI is running.
• Model lists are fetched from the local Ollama server.
• Stored runs are saved as numbered JSON files.