🎨 Polza Text‑to‑Image

Polza Text-to-Image is the node that turns your graph into a paying OpenAI customer. It's a thin wrapper over Polza.ai's OpenAI-compatible image endpoint - /v2/images/generations - and it exposes exactly the three models that endpoint knows: gpt-image-1 (GPT Image), dall-e-3, and dall-e-2. None of these have open weights, so you can't run them locally, full stop. If you want a GPT Image render inside your workflow, this is the door - the "call the closed model" path rather than the "download the checkpoint" path.

Why reach for it specifically? gpt-image-1 is genuinely good at text rendering and following fussy instructions, which is exactly what local checkpoints still stumble on. And because it's OpenAI-compatible, the node slots into the same patterns the rest of the pack uses: chain Polza Chat into it for a write-the-prompt-then-render pipeline, or describe an image with Polza Vision and feed its text into this node's prompt for a describe-and-redraw loop.

How it works

The node builds the generation request with a response_format: url, calls the API, and - because these models can take a while - polls until the job completes, then downloads the returned images and stacks them into a standard IMAGE tensor. One detail worth knowing: it's model-aware. DALL·E 3 gets the style param; gpt-image-1 gets output_format, background, and output_compression; and DALL·E 3's revised_prompt is surfaced as an output rather than hidden.

The inputs that matter

• model - gpt-image-1 (default), dall-e-3, dall-e-2. Click Load models on the node to populate the dropdown - it won't fetch until you do.
• prompt - required; up to 32,000 characters for gpt-image-1, so it can swallow a long LLM-generated description whole.
• size - auto plus the usual list (1024x1024, 1792x1024, 1024x1792, …). Default is 1024x1024.
• quality - auto/high/medium/low/hd/standard.
• style - vivid/natural; only applied to DALL·E 3.
• n - how many images. DALL·E 3 accepts only 1, so leave it alone unless you're on gpt-image-1.

Outputs

Three: images (the IMAGE tensor - wire it to a Preview Image or Save Image node like any generator), revised_prompt (STRING - DALL·E 3 rewrites your prompt internally and reports the rewrite here; gpt-image-1 keeps it empty), and cost_rub (what the call cost, in rubles).

Installing it

It ships with the ComfyUI-Polza pack. ComfyUI Manager → search ComfyUI-Polza → install → restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/PioneerMNDR/ComfyUI-Polza
cd ComfyUI-Polza
pip install -r requirements.txt

Restart, grab a key from polza.ai, and set it via the node's api_key field, the POLZA_API_KEY env var, or a config.json in the node folder. No model downloads - the models live on OpenAI's side.

Common issues

The revised_prompt output is the first thing that confuses people. DALL·E 3 doesn't generate from what you typed; it silently rewrites your prompt into something it considers safer/more complete, and generates from that. If your output doesn't match your intent, read the revised prompt - that's your answer. It also means your prompt is getting laundered through OpenAI's moderation regardless of what Polza's own enable_safety-style settings suggest.

Other gotchas: DALL·E 3 ignores n > 1 and the gpt-image-1-only params do nothing on DALL·E 2. And the economics you already know - every queue run is a paid call, cost in rubles, data (including your prompt) leaves the machine, and OpenAI's content policy is the ceiling on what this node can produce. The code is MIT-licensed and short enough to read before you trust it. For a text-heavy poster you can't do locally, it's worth it; for a generic landscape you could Flux at home for free, you're just paying for the privilege of the API.