Model-specific options¶

Gemini can return a whole set in one request with --mode batch.
Gemini 3.1 can think harder before it draws with --thinking high.
Gemini can reply with text. genimg prints it but does not save it.
Only GPT Image 2.5 has xhigh and max quality.
Codex picks the model and size for you. --aspect-ratio is only a request.

--num, --diverse, --deltas, --output and --grid work with every model. genimg rejects a flag the model can't use, even on --dry-run.

Flag	Gemini (`gdm:`)	OpenAI (`oai:`)	Codex (`codex:image`)
`--mode batch`	✅	❌	❌
`--thinking`	✅ `gdm:nb2`, `gdm:nb2-lite` only	❌	❌
`--quality`	❌	✅ `xhigh`, `max` on 2.5 only	❌
`--resolution`	✅ by model	✅ size table, `1K` only on GPT Image 1.x	❌
`--aspect-ratio`	✅ 10 ratios, 14 on `gdm:nb2`, `gdm:nb2-lite`	✅ 5 ratios, `1:1` only on GPT Image 1.x	✅ as a prompt request
`--input`, reference images	✅	✅ up to 16 files	✅
`--region`	✅	❌	❌
`--project`	✅	❌	❌
`--auth`	❌	✅ `azure` or `direct`	❌
Text reply	✅ printed, not saved	❌	❌ dropped

Gemini¶

With --mode batch --diverse, the model varies its own takes. See Diverse images.
A batch can come back short. The model picks the count. genimg keeps what arrives and warns.
--thinking high costs more than the estimate. Google bills thinking tokens, and genimg leaves them out.
Only text before the image is printed. genimg drops the rest, and all text in batch mode.

$ genimg "a minimal fox logo" \
    --model gdm:nb2 \
    --num 4 \
    --diverse \
    --mode batch \
    --thinking high \
    --dry-run
genimg google/direct gdm:nb2 → gemini-3.1-flash-image
  prompt   "a minimal fox logo"
  params   n=4 mode=batch diverse thinking=high
  cost     $0.2680 (estimate)  id=20260925_115942_6af52e
  outputs  ~/.genimg/generations/20260925_115942_6af52e_1.png
           …
  diverse  model-coordinated: the single batched request asks for deliberately different takes
dry-run: no API call made.

OpenAI¶

--quality defaults to medium. GPT Image 2.5 adds xhigh and max.
--num sends one request per image. A batch returns near-duplicates, so genimg rejects it.
GPT Image 1, 1.5 and 1 mini take 1024×1024 only.
Up to 16 input and reference images. PNG, JPEG or WebP, 50 MB each.

$ genimg "a lighthouse at dusk" \
    --model oai:gi2.5 \
    --quality xhigh \
    --resolution 2K \
    --aspect-ratio 16:9 \
    --dry-run
genimg openai/direct oai:gpt-image-2.5-sunburst → gpt-image-2.5-sunburst
  prompt   "a lighthouse at dusk"
  params   n=1 q=xhigh r=2K a=16:9 → 2048x1152
  cost     $0.0753 (estimate)  id=20260925_115943_ca5e9a
  output   ~/.genimg/generations/20260925_115943_ca5e9a.png
dry-run: no API call made.

Codex¶

--aspect-ratio is only a request. genimg adds it to the prompt, so check the result.
Runs use your Codex allowance. genimg's API-equivalent price is for comparison, not a charge. See Codex subscription.

$ genimg "a lighthouse at dusk" \
    --model codex:image \
    --aspect-ratio 16:9 \
    --dry-run
genimg codex/subscription codex:image → codex:image
  prompt   "a lighthouse at dusk"
  params   n=1 a=16:9
  cost     Codex subscription (usage limits apply)  id=20260925_115944_aa2c84
  runtime  Codex subscription selects the image model and size; aspect ratio is a prompt request.
  output   ~/.genimg/generations/20260925_115944_aa2c84.png
dry-run: no API call made.