Image generation

Pendra generates images from text prompts using diffusion models, served through the OpenAI-compatible images endpoint. Install an image model once and you can generate from the API, the SDKs, or the Playground.

Generate an image

Send a model and a prompt. Results come back as base64-encoded PNGs in data[].b64_json — decode and write them to disk.

from pendra import Pendra
import base64, pathlib

client = Pendra()
result = client.images.generations.create(
    model="flux.1-schnell",
    prompt="A misty Welsh valley at dawn, painted in oils",
)
pathlib.Path("out.png").write_bytes(base64.b64decode(result.data[0].b64_json))

import Pendra from 'pendra';
import { writeFileSync } from 'node:fs';

const client = new Pendra();
const result = await client.images.generations.create({
  model: 'flux.1-schnell',
  prompt: 'A misty Welsh valley at dawn, painted in oils',
});
writeFileSync('out.png', Buffer.from(result.data[0].b64_json, 'base64'));

Response

Available models

Several text-to-image models are one install away — for example:

• sdxl-turbo — fast 1–4 step generation.
• flux.1-schnell — FLUX.1 [schnell], high quality (~10 GB).
• stable-diffusion-3.5-large-turbo — SD 3.5 Large Turbo.
• z-image-turbo — Z-Image Turbo, fast few-step generation.
• flux.2-klein — FLUX.2 [klein], in 4B and 9B sizes.
• qwen-image-2512 — Qwen-Image 2512, strong at rendering text in images.

Browse what's available at /models?type=image.

Options

By default Pendra generates at each model's native resolution — the size it was trained for, which gives the best quality, so you usually don't need to set size at all. That's 512x512 for sdxl-turbo and flux.2-klein, and 1024x1024 for flux.1-schnell, z-image-turbo, qwen-image-2512, and stable-diffusion-3.5-large-turbo. You can still request a specific size (WxH) when you need a particular shape — but note that generating far from a model's native resolution can distort the result (for example, duplicated subjects).

Use num_inference_steps to trade speed for quality: more steps generally means a sharper, more coherent image but a slower generation. Like size, it defaults to each model's recommended step count — the same number the console Playground's Steps slider starts at, so a request from your code and one from the Playground render alike. That is just 4 for sdxl-turbo, flux.1-schnell, stable-diffusion-3.5-large-turbo and flux.2-klein, which are built to converge in a handful of steps, 8 for z-image-turbo and krea-2-turbo, and 20 for qwen-image-2512 — so you usually don't need to set it. It accepts values from 1 to 50. The same defaults apply with private inference: Pendra can't read an encrypted request, so your worker fills them in. Keep your workers up to date, as an older worker falls back to 4 steps on the "turbo" / "schnell" models and 30 on the rest.

Guidance strength — how hard the model is pushed towards your prompt — is chosen per model too, and there is no parameter for it. The distilled few-step models (the "turbo" and "schnell" models and flux.2-klein) already have that guidance trained in, so Pendra samples them gently; pushing them as hard as a standard checkpoint is what makes their output look noisy and over-saturated. Standard checkpoints get the stronger setting they expect.

Also tune n (up to 4 images), negative_prompt, and seed for reproducibility. Generation is slower than chat — expect a few to thirty seconds depending on model, size, and steps — so render a spinner in interactive UIs.