Grok Text Refine (xAI, legacy)

What it is

Grok Text Refine takes your text, sends it to xAI's chat API alongside a system_prompt, and gives you back whatever the system prompt asked for. That's it. "Refine" is the author's use case, not a constraint - the same node is a caption refiner, a prompt generator, a translator or a summarizer, and which one you get is entirely decided by the text you put in system_prompt.

If you've seen the "LLM in a graph" pattern, this is it - and note what it is not. It is not the text encoder inside your checkpoint, which is frozen, part of the model, and runs every generation. This is a separate language model bolted onto the canvas as a tool, working upstream of the sampler. The two get conflated constantly. The community settled on the pattern a while ago - "prompt enhancer" mentions went from 12 in 2023 to 253 in the first half of 2026 - and the two shapes it takes here are:

[Load Image] → [JoyCaption] → [Grok Text Refine (xAI)] → [Save Text File]

[Grok Text File Loader] → [Grok Text Refine (xAI)] → [CLIP Text Encode] → [KSampler]

The first is deterministic caption cleanup. The second is prompt generation with real variety, one idea per queue run. Keep that split in mind, because it decides your seed setting later.

How it works

It's an OpenAI-compatible POST to https://api.x.ai/v1/chat/completions - your text as the user message, your system prompt as the system message - and it does a decent amount of housekeeping you'd otherwise be writing yourself.

It retries: three attempts total, with backoff, for timeouts and the usual transient HTTP codes. It cleans the reply before handing it to your graph, stripping stray markdown fences and wrapping quotes that chat models love to add. And it handles the reasoning-model parameter problem in two ways. Neutral sampling values (top_p at 1.0, penalties at 0.0) are never sent to the API at all - which is why models like grok-4.5 that reject presencePenalty outright don't break. If a model still rejects one, the node reads the error, strips that parameter, and resends without it, logging which one it dropped. The queue keeps running.

The inputs that matter

• system_prompt - the job. This is the one you'll spend your time on. Feed it from Grok System Prompt Loader so you can edit it in a text editor instead of a widget.
• text - the user message. A raw caption to refine, or a short idea to expand.
• temperature - 0.2 by default, which is right for caption refinement. For prompt generation push it to 1.0–1.15. Above ~1.3, grammar starts to go.
• seed - not sent to the API. It's purely a cache-breaker. ComfyUI only re-runs a node when an input changed, so with fixed and identical inputs you get the cached result back and no API call. Set it to randomize for prompt generation, or you'll generate every image from literally the same prompt text. Leave it fixed for caption refinement and repeated runs cost you nothing.
• model - defaults to grok-4. Model ids change; check curl https://api.x.ai/v1/models -H "Authorization: Bearer YOUR_KEY" rather than trusting the default.
• fallback_to_input - on by default. If the call fails, your input text passes through unchanged instead of killing the queue. Leave it on for batch work.

Output is a single refined_text string. Wire it to CLIP Text Encode for a seed prompt, or to Save Text File for dataset work.

Install

Manager: search Grok Prompt Forge (xAI). Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/Slartibart23/comfyui-grok-prompt-forge

Restart; nodes land under Grok → Text. This pack is the successor to comfyui-grok-text-refine with the same node class IDs, so old workflows load unchanged - but delete the old comfyui-grok-text-refine folder when you upgrade, or you'll register the same class twice.

Get a key at console.x.ai and put it in grok_api_key.txt in your ComfyUI root with the api_key widget empty - a key typed there gets saved into any workflow you export.

Common issues

Replace the default system prompt. The one baked into this node is the author's LoRA dataset-captioning prompt, complete with a specific trigger token and rules about what to strip from captions. Connect a text source and wonder why your prompts come back looking like training captions - that's why. Good template, not a neutral default.

The penalty dials can silently do nothing. On reasoning models that reject them, the node strips and retries - so the run succeeds and the console is the only place you'll see that your presence_penalty had no effect. On those models you get variety from temperature plus an explicit ban list in the system prompt, and that ceiling is worth knowing anyway: sampling works at the word level, and it can't stop the model loving the same motifs - the same centered poses, the same lighting clichés. Ban clauses and genuinely different input text (see Grok Text File Loader cycling a folder) are the fix.