Prompting

Trust labels: [vendor] official model-author doc, [measured] measured on DeepInfra by the webtoon project (2026-09), [community], [inference].

Structure (all image models)

1. Order is weight. Subject + action, then style/medium, then setting, lighting, camera/technical. FLUX prioritises earlier words [vendor: BFL]. A composition instruction placed after 400 characters of style text was ignored; moved to the front it took effect [measured].
2. Prose, not keyword lists. [vendor: BFL, Google, Qwen]. The bad example from BFL: woman, portrait, beautiful, blonde, studio, 8k.
3. Always one lighting sentence. "It has the single greatest impact on image quality" [vendor: BFL].
4. Name the medium ("flat vector illustration", "gouache", "cel animation"). Medium moves the image more than mood [measured]. Naming 2-3 specific colours beats a vibe; the line description is the strongest style lever [measured].
5. Be concrete. Materials, colours with a modifier, counts as words. Enumerate, never summarise [vendor: Qwen rewriter].
6. State the use ("app icon", "landing hero") to set polish [vendor: OpenAI].
7. No quality boosters ("8k", "masterpiece", "award-winning") [vendor: Qwen]; generic negative dumps degrade distilled models [community].
8. Aspect ratio and size are parameters, never prose [vendor: Qwen].
9. Length. 30-80 words for most work; 10-30 for concepts; 80+ for complex scenes on FLUX.2 (its encoder is a 24B VLM) [vendor: BFL]. Schnell's encoder takes ~256 tokens: keep to 40-75 words [community]. Qwen-Image likes 400-500 word observer-style descriptions [vendor].
10. Every word in a stock description ends up in every image. A prop in a location string appeared in most renders [measured]. Keep constant strings to what is always visible.

Per-family differences

Family	Length	Negative prompt	Notes
FLUX-1-schnell	40-75 words	none; phrase positively	guidance 0, 1-4 steps, no upsampling: the prompt must carry the detail [vendor]
FLUX-2 dev/klein/pro/max	30-80, up to several hundred	none	JSON prompts, hex colours tied to objects, real camera/film names work [vendor]; multi-reference as "image 1 / image 2"
Gemini image (nano-banana 2 / pro)	conversational, purpose stated	none; "semantic negatives"	start with a strong verb; last input image sets the ratio; native route on DeepInfra takes ONE reference [measured]
Qwen-Image / -Max	long, descriptive	field exists; leave " "	append ", Ultra HD, 4K, cinematic composition." [vendor]; prompt_extend defaults on [vendor: DeepInfra schema]
Qwen-Image-Edit	imperative	field exists	<image1> role tags [vendor]
Seedream 5	prose	none	optimize_prompt_mode rewrites for you; background: transparent supported
Wan T2I	prose	field exists	prompt_extend defaults on

Positive substitutions (never write "no X" on FLUX, Gemini, Qwen) [vendor: BFL]

Instead of	Write
no people	empty, deserted, solitary
no text / no watermark	clean surfaces, unmarked, text-free, pristine
no logos	unbranded, plain, logo-free surface
no blur	sharp focus, crisp details, tack-sharp
no CGI look	photorealistic, authentic
no clouds	clear blue sky, cloudless
no background	isolated on a plain solid white background
no hat	bare head

Negative phrasing "can actually make models focus MORE on unwanted elements" [vendor: BFL]. On routes with no negative field vincevangen folds --negative into the prompt as "Avoid: …"; that helps on Gemini-family models and is neutral-to-harmful on FLUX [inference], so prefer positive phrasing there.

Hex colours [vendor: BFL]

Tie a hex to an object: "walls in #C4725A", "gradient from #02eb3c to #edfa3c". Never "use red somewhere". Only use hex when the brief gives it; otherwise a modifier + colour name ("deep navy") [vendor: Qwen].

Typography [vendor: BFL, Google]

• Quote the exact string; spell unusual words letter by letter.
• Name the font class (serif, sans-serif, script, display, monospace), weight, colour, size hierarchy, placement ("upper third", "bottom right corner").
• Front-load: A sign reading "FRESH BREAD" in a bakery window beats the reverse.
• Lay out multi-line text line by line. Keep to 1-3 short strings.
• Generate art text-free and letter in post whenever the copy may change: rewriting a line must never cost a render [measured: working practice]. vincevangen's post step is text_overlay / vvgen_add_text (bundled OFL fonts, free).
• 2026-10-09, one seed each, "LAUNCH" on an icon: FLUX-2-dev wrote "LAUCH"; klein-4b and pro spelled it right [measured]. Model-painted words are a draw, so check them.
• The text-free instruction still fails sometimes; gate with OCR or by eye.

JSON prompts (FLUX.2) [vendor: BFL]

{"scene": …, "subjects": [{"description": …, "position": …, "color": …}], "style": …, "lighting": …} works for multi-subject layouts where each subject needs a position.

The enhancer's rules (what --enhance does)

Adapted from the Qwen-Image-2.1 T2I rewriter and OpenAI's augmentation policy: split the brief into fixed (copied verbatim) and open (invented tastefully); add nothing unrequested (no extra characters, props, brands, slogans); normalise an already-specific brief instead of expanding it; opening sentence with medium noun; walk the frame; one lighting sentence; closing "The overall composition…"; honour a word budget per model family; output JSON with prompt, aspect_ratio, avoid (positive). Ratio defaults: 3:2 horizontal, 2:3 vertical, 1:1 for a single centred emblem, 16:9 cinematic/slide, 9:16 phone.