Skip to content

Text to Image

Generate a PNG from a text prompt. The cheapest and most mature task to run end-to-end on OmniRT.

Minimal example

from omnirt import generate
from omnirt.requests import text2image

result = generate(text2image(
    model="sd15",
    prompt="a lighthouse in fog, cinematic, 35mm film",
    preset="fast",
))
print(result.artifacts[0].path)

omnirt generate \
  --task text2image \
  --model sd15 \
  --prompt "a lighthouse in fog, cinematic, 35mm film" \
  --preset fast \
  --backend auto

# request.yaml
task: text2image
model: flux2.dev
backend: auto
inputs:
  prompt: "a cinematic sci-fi city at sunrise"
config:
  preset: balanced
  width: 1024
  height: 1024

omnirt generate --config request.yaml --json

curl -sS http://localhost:8000/v1/generate \
  -H 'Content-Type: application/json' \
  -d '{
    "task": "text2image",
    "model": "sd15",
    "inputs": {"prompt": "a lighthouse in fog"},
    "config": {"preset": "fast"}
  }'

Key parameters

Parameter Type Default Notes
prompt str required text prompt
negative_prompt str? None negative prompt; honored by SD / SDXL / SD3
width / height int? model default output size; models enforce 8/16/32-multiple constraints
preset fast / balanced / quality / low-vram balanced bundled steps / precision / guidance; see Presets
num_inference_steps int? preset explicit denoise step override
guidance_scale float? preset classifier-free guidance
num_images_per_prompt int? 1 batch images per prompt
seed int? random fix randomness for reproducibility
scheduler str? model default see Architecture → Scheduler layer
dtype fp16 / bf16 / fp32 fp16 compute dtype; Ascend defaults to bf16
adapters list[AdapterRef]? [] LoRA / ControlNet adapters

Supported models

Typical quality / speed tradeoffs:

  • Highest quality: flux2.dev (≥ 24 GB VRAM), sdxl-base-1.0 + sdxl-refiner-1.0
  • Balanced: sdxl-base-1.0, sd3-medium, qwen-image
  • Low-resource: sd15 (12 GB OK), sd21

Full list: omnirt models or Supported Models.

Common recipes

omnirt generate --task text2image --model sd15 \
  --prompt "..." --preset fast --backend cuda

omnirt generate --task text2image --model flux2.dev \
  --prompt "..." --preset quality --width 1024 --height 1024 \
  --backend cuda --dtype bf16

omnirt generate --task text2image --model sd15 \
  --prompt "..." --preset low-vram --dtype fp16

Troubleshooting

Common issues

  • ValidationError: width must be multiple of 8 — most SD-family models require multiples of 8; Flux2 is stricter (16 / 32)
  • CUDA out of memory — switch to --preset low-vram or reduce width/height; or set OMNIRT_DISABLE_COMPILE=1 to skip torch.compile
  • adapter not supported for this model — check omnirt models <model_id>'s adapters field; LoRA / ControlNet compatibility is declared in ModelCapabilities

Running omnirt validate catches the first two without touching a GPU.