Image Gen
The first generation should be close. Provider choice and prompt structure
decide that, not luck.
Provider routing
| Job |
Provider |
Why |
| Text in image (posters, UI mockups, signage) |
gpt-image (gpt-image-2) |
Best-in-class verbatim text rendering |
| Design fidelity, brand-adjacent work |
gpt-image |
Follows structured art direction |
| Fast iteration, photoreal people |
grok (via groken cloud) |
Speed, identity realism |
| Textures, tiles, sprites |
gpt-image low quality first, then high |
Cheap drafts before the final render |
Check what's installed before promising: the harness's generate_image
tool for gpt-image; the groken bridge for grok. Never fake a provider that
isn't there — say what's missing.
Prompt structure (the part people skip)
- Subject — one clear subject, named concretely ("a ceramic pour-over
coffee dripper" not "coffee thing").
- Composition — where the subject sits, camera angle, framing
("centered, slight low angle, product shot on seamless").
- Style — the visual language ("editorial product photography",
"flat vector illustration", "35mm film grain").
- Lighting — the single biggest quality lever ("soft key from upper
left, subtle rim light", "overcast window light").
- Negatives — what to avoid ("no text, no watermark, no hands").
- Verbatim text — if text appears, put it in quotes exactly
(
the words "SALE ENDS FRIDAY") and keep it short.
Iteration discipline
Draft cheap (low quality), pick a direction, then render the final at high
quality. Read the model's revised_prompt when available — it tells you what
it actually heard. One strong revision beats five blind regenerations.
Pairs with
design (the taste standard the image is judged against),
higgsfield-generate (heavier video/3D/audio generation when installed).
1---2name: image-gen3description: Image Gen4---56# Image Gen78The first generation should be close. Provider choice and prompt structure9decide that, not luck.1011## Provider routing1213| Job | Provider | Why |14|---|---|---|15| Text in image (posters, UI mockups, signage) | gpt-image (gpt-image-2) | Best-in-class verbatim text rendering |16| Design fidelity, brand-adjacent work | gpt-image | Follows structured art direction |17| Fast iteration, photoreal people | grok (via groken cloud) | Speed, identity realism |18| Textures, tiles, sprites | gpt-image low quality first, then high | Cheap drafts before the final render |1920Check what's installed before promising: the harness's `generate_image`21tool for gpt-image; the groken bridge for grok. Never fake a provider that22isn't there — say what's missing.2324## Prompt structure (the part people skip)25261. **Subject** — one clear subject, named concretely ("a ceramic pour-over27 coffee dripper" not "coffee thing").282. **Composition** — where the subject sits, camera angle, framing29 ("centered, slight low angle, product shot on seamless").303. **Style** — the visual language ("editorial product photography",31 "flat vector illustration", "35mm film grain").324. **Lighting** — the single biggest quality lever ("soft key from upper33 left, subtle rim light", "overcast window light").345. **Negatives** — what to avoid ("no text, no watermark, no hands").356. **Verbatim text** — if text appears, put it in quotes exactly36 (`the words "SALE ENDS FRIDAY"`) and keep it short.3738## Iteration discipline3940Draft cheap (low quality), pick a direction, then render the final at high41quality. Read the model's revised_prompt when available — it tells you what42it actually heard. One strong revision beats five blind regenerations.4344## Pairs with4546`design` (the taste standard the image is judged against),47`higgsfield-generate` (heavier video/3D/audio generation when installed).
trac3r00