On this page

Image Generation

The image-generation node generates a new image with a diffusion model. Unlike the analysis nodes, which describe a property of the media itself, imagegen produces a fresh image — either from a text prompt (GENERATE) or by remixing the asset’s own image guided by a prompt (REMIX) — and attaches it to the asset.

The diffusion model runs behind a small HTTP sidecar, so the node stays a pure client and needs no model runtime of its own. The sidecar is model-agnostic: it ships with SDXL-Turbo by default (ungated, fast) and can serve other diffusers-compatible models.

Kind

imagegen

Applies to

Image assets

Inputs

A generation prompt from the node configuration; in REMIX mode also the asset’s own image

Output keys

imagegen_flag (String), imagegen_path (String, local path of the generated PNG)

Requirements

A running image sidecar (/generate, /remix). A GPU is recommended for interactive latency.

Persists to

asset_node_result ledger only — the generated PNG stays in the worker’s local imagegen_bin cache (as with thumbnails)

Configuration

ImageGenNodeOptions:

Option Meaning

mode

GENERATE (text-to-image) or REMIX (image-to-image) (default GENERATE)

prompt

The generation prompt (required)

host / port

Address of the image sidecar (default localhost:9200)

width / height

Output size for GENERATE (default 1024 × 1024)

strength

Denoise strength for REMIX, in (0, 1] (default 0.6)

steps

Number of inference steps (default 30)

seed

Optional RNG seed for reproducible output

Running the sidecar

The sidecar is a small FastAPI service. A minimal run with the default (ungated) model:

cd sidecars/ideogram-sidecar
python -m venv venv && ./venv/bin/pip install -r requirements.txt
./venv/bin/uvicorn server:app --host 0.0.0.0 --port 9200
# smoke test
curl -s -X POST localhost:9200/generate \
  -H 'content-type: application/json' \
  -d '{"prompt":"a red panda astronaut, studio lighting"}' -o out.png

Use Cases

  • Illustrations & hero images — generate on-brand artwork from a prompt for a catalogue entry.

  • Restyling — remix an existing asset into a new style while preserving its composition.

  • Placeholders & variations — produce derived visuals for review workflows.

The generated PNG is written to the worker’s local imagegen_bin cache and a node-result ledger row records that the node ran for the asset.