Presets

omnirt ships four out-of-the-box presets: fast, balanced, quality, low-vram. Each preset is a small set of config overrides. The request's config is merged in this order, with later layers winning:

1. Base preset — steps / dtype shared by all tasks
2. Task preset — per-task guidance-scale defaults (text2image / text2video / image2video)
3. Model preset — narrow model-specific overrides (today, the Flux2 family)
4. User-provided config — always wins

The merge happens in resolve_preset(task, model, preset) in src/omnirt/core/presets.py.

Base preset

Preset	num_inference_steps	dtype
fast	20	—
balanced	—	—
quality	40	—
low-vram	18	fp16

balanced keeps each pipeline's own default step count. low-vram forces 18 steps and fp16 so memory-constrained devices still finish.

Task preset — guidance_scale

Preset	text2image	text2video	image2video
fast	5.5	4.0	3.0
balanced	7.5	5.0	3.5
quality	8.0	6.0	4.0
low-vram	—	4.0	3.0

image2image / inpaint / edit reuse the text2image entries (see _TASK_PRESET_ALIASES). low-vram on text2image only contributes the base preset's steps + dtype and does not override guidance.

Model preset — flux2.dev / flux2-dev

Preset	guidance_scale	max_sequence_length	Other
fast	2.0	384	—
balanced	2.5	512	—
quality	3.0	512	—
low-vram	2.0	256	dtype=fp16

Flux is sensitive to high guidance; these values override the task preset to stay in the useful range.

Usage

omnirt generate \
  --task text2image \
  --model sd15 \
  --prompt "a lighthouse in fog" \
  --preset fast

Python equivalent:

from omnirt import requests, generate

req = requests.text2image(model="sd15", prompt="a lighthouse in fog", preset="fast")
result = generate(req, backend="cuda")

Inspecting the merged result

omnirt validate \
  --task text2image --model flux2.dev \
  --prompt "..." \
  --preset low-vram \
  --backend cpu-stub

The validator returns a resolved_config that shows exactly what the preset produced, so you can see how the merge resolved for the specific model you chose.