GROK TEXT REFINE — SETTINGS GUIDE ================================== WHAT THIS NODE DOES Sends your input text plus a system prompt to the xAI Grok API and returns the generated text. Two typical modes: A) CAPTION REFINEMENT -> you want the SAME output every time B) PROMPT GENERATION -> you want a FRESH output every time ---------------------------------------------------------------- THE OPTIONS ---------------------------------------------------------------- model Which Grok model handles the request (e.g. grok-4.5). Note: some reasoning models partially ignore sampling settings. temperature (0.0 - 2.0) The main creativity dial. It reshapes the probability of every word choice: low values make the model always pick its favorite words (predictable, repetitive), high values give unusual words a real chance (varied, surprising). 0.0 - 0.3 -> near-deterministic. Use for caption refinement. 0.7 - 0.9 -> balanced variety. 1.0 - 1.15 -> high variety. Use for prompt generation. > 1.3 -> risk of broken grammar and word salad. top_p (0.0 - 1.0) A cutoff filter. Before picking a word, the model sorts all candidates by probability and THROWS AWAY the unlikely tail. top_p = 0.9 keeps only words covering the top 90% of probability; everything below is impossible to pick, no matter the temperature. 1.0 -> nothing is cut off (full variety possible). 0.95-0.98 -> trims only true nonsense. Good default. < 0.9 -> noticeably safer/blander output. RULE OF THUMB: temperature shapes the dice, top_p removes dice faces entirely. For variety: keep top_p HIGH (0.98-1.0) and steer with temperature. presence_penalty (-2.0 - 2.0) Pushes the model away from words it has ALREADY USED in this output, regardless of how often. Great against the model's favorite stock phrases ("golden hour", "neon lights"...). 0.0 -> off (default). 0.3 - 0.6 -> gentle push toward new vocabulary. Recommended for prompt generation. > 1.0 -> can get incoherent. frequency_penalty (-2.0 - 2.0) Like presence_penalty, but the push grows THE MORE OFTEN a word repeats. Targets literal word repetition inside one output. 0.0 -> off (default). 0.2 - 0.5 -> reduces "she... she... she..." loops. seed + control_after_generate NOT sent to the API. This is a cache-breaker: ComfyUI only re-runs a node when one of its inputs changed. randomize -> node fires a NEW API call every queue run. Required for prompt generation! fixed -> identical requests reuse the cached result (saves money on repeated caption runs). max_tokens Upper limit for the response length. 1024 is plenty for a 150-200 word image prompt. Too low = truncated prompts. timeout_seconds / fallback_to_input Network safety: how long to wait, and whether to pass the input text through unchanged if the API call fails (so a batch never crashes mid-run). ---------------------------------------------------------------- RECOMMENDED PRESETS ---------------------------------------------------------------- PROMPT GENERATION (fresh, varied prompts): temperature 1.0-1.1 | top_p 0.98 | presence_penalty 0.4 frequency_penalty 0.2 | seed: randomize CAPTION REFINEMENT (stable, repeatable): temperature 0.2 | top_p 1.0 | penalties 0.0 | seed: fixed STILL GETTING SIMILAR IMAGES? Sampling settings add word-level variety, but they cannot stop the model from loving the same MOTIFS (same poses, same golden- hour backdrops). Fix that in the system prompt: add an explicit ban list ("Never use: golden hour, neon, rain-slicked streets, subject centered facing camera...") and/or feed randomized scene ingredients (location, camera angle, pose, light) into the CONTENT line via wildcards.