Prompt Benchmark Rating

The honest problem with LLM-assisted prompting is that your feedback to the model is usually a shrug: "looks off, make it better." An LLM can't do much with that. PBRating exists to turn your gut feeling into structured, calibrated feedback the Prompt Bench text nodes actually use - you score six criteria on a 1–5 scale, and it hands the LLM a tidy package of what to fix, in what order.

What it is

It's the rating end of the loop from inflamously/comfyui_prompt_bench. You look at the image the sampler just produced, score it, and out pops a single rating payload. That payload feeds straight into PBTextModel or PBTextModelChat (and PBStore, if you want to log it). The node doesn't touch Ollama and doesn't need a GPU - it's just a structured form with a little math behind it.

The criteria that matter

Six drop-downs, each 1 - Poor through 5 - Excellent, defaulting to 3 - Okay, plus a freeform notes field. You set all of them; they're the whole node:

• prompt_adherence - does the image match what you actually typed?
• aesthetics - color harmony, composition, overall appeal
• technical_quality - sharpness, artifacts, clean render
• lighting - light quality, shadows, highlights, mood
• detail_richness - surface textures, fine detail
• coherence - anatomy, physics, spatial sense

Output: rating (RATING) - a dict with a rating_id, your notes, integer scores per criterion, and an automatically computed average_score.

How the score becomes instruction

This is the part that makes it better than freeform critique. The pack's prompt.py turns each score into a concrete rewrite rule: a 1 on lighting becomes "explicitly define a single primary light source with exact position," a 1 on coherence becomes "enforce a physically consistent layout." Scores of 4–5 get "maintain, only minor precision improvements." It even orders the feedback - fix the lowest scores first, then apply your notes. Your one-line note ("too busy background") gets appended as a hard requirement the model is told to apply precisely.

The one habit that matters: don't leave everything at 3. Three scores produce mild nudges ("refine lighting: specify intensity"), which is what you want mid-iteration, but if you're genuinely unhappy, mark a 1 or 2 - that's the signal that actually changes the prompt. And since the payload is computed per run, any edit to a score means re-executing the node; there's no "live update."

Install and wiring

Same as the rest of the pack - ComfyUI Manager, search "Prompt Bench", or:

cd ComfyUI/custom_nodes
git clone https://github.com/inflamously/comfyui_prompt_bench

restart, done. No pip install, no models to download. Wire rating into PBTextModel / PBTextModelChat and optionally into PBStore for the audit trail.

Gotchas

• The RATING output type is pack-specific - only the other Prompt Bench nodes (and anything else that declares it) accept it.
• notes is optional but worth using; it's the only place your subjective take ("the hands look wrong") survives the rubric.

It's a small node with a small job: make your feedback legible to a local LLM. Once you've scored a few images with it, the 1–5 rubric quietly becomes the way you look at every generation.