Cosmos 3 Super Prompting Guide

Cosmos 3 Super is NVIDIA's text-to-image model on Versely. This page is its structured prompting reference: the 5 parameters its schema actually exposes, the t2i technique that applies to it, copy-ready templates.

Everything here is grounded in the same sources Versely's agent reads — the model's input schema. Where a line is general craft advice rather than a documented fact about Cosmos 3 Super, the page says so.

What Cosmos 3 Super wants

The exact input surface, from the same schema the Versely agent fetches with get_model_input_schema before every generation.

Technique that applies here

Text-to-image: composition, lighting, style; family tips (Midjourney/Flux/Imagen/etc.) drive syntax

• Qwen Image (+ Qwen Z Image), GPT Image 2, Nano Banana, Cosmos 3 Super, Krea 2 Medium, Grok Imagine Image, Recraft 4.1, and ImagineArt 2.0 aren't in that 12-family list, so they fall back to the enhancer's generic default: 'vivid, specific descriptions with details about composition, lighting, style, and atmosphere.' No model-specific syntax is being layered on for you on these — be explicit yourself about subject, setting, lighting, style, and mood.
• Pick aspect ratio / size from the model's actual enum rather than describing it in prose. Cosmos 3 Super and Recraft 4.1 both take the same six image_size presets (square_hd, square, portrait_4_3, portrait_16_9, landscape_4_3, landscape_16_9); Flux 2 Max's resolved schema takes that same six plus a 7th value, auto; Seedream 4.0 uses the six-preset list too, plus a custom {width, height} object (960-4096px, ~2048x2048 default); Grok Imagine Image Quality and Krea 2 Medium instead take an aspect_ratio string enum (Krea's list includes a 2.35:1 cinematic option Grok's doesn't have).
• Use a dedicated schema field instead of writing it into prose when one exists. Cosmos 3 Super has a real negative_prompt field for exclusions (separate from the main prompt). Krea 2 Medium has image_style_references — attach style image(s) there to carry a look instead of spelling it out in adjectives. Seedream 4.0 has its own enhance_prompt_mode (standard/fast) built into the provider API — that's a separate toggle from Versely's own AI-enhance step, not the same thing.
• Where the schema exposes num_images / max_images (Cosmos 3 Super and Grok Imagine Image Quality up to 4, Seedream 4.0 up to 6), request a batch instead of one-at-a-time generations — compare compositions in a single call before you commit a direction to a full prompt rewrite.

Copy-ready templates

Replace the bracketed slots; each template says when it's the right shape.

Prompt: [SUBJECT], [SETTING], [STYLE], [LIGHTING]  //  Negative prompt (separate field, not inline text): [UNWANTED ELEMENTS, e.g. blurry, extra limbs, text, watermark]

[SUBJECT], [SETTING], [STYLE] — set num_images/max_images to [2-4] and generate variants in one call before refining the prompt further

Mistakes that waste generations

• Padding every prompt with quality-booster tag lists ("masterpiece, best quality, highly detailed, 8k"). That convention belongs to the 'stable' MODEL_TIPS family — comma-separated tags plus quality boosters — and nothing in this family resolves to it; on Flux, Imagen, Midjourney, or any of the fallback-tip models it just reads as filler the model has to parse around.
• Free-typing an aspect ratio or size string outside the model's real enum — e.g. asking for 'a wide 2.5:1 banner' on a model whose aspect_ratio list tops out at 21:9, or writing 'widescreen' where the field wants the literal preset value landscape_16_9. The schema enum is exact; the model isn't interpreting loose ratio language.
• Writing an edit instruction ("make the sky sunset-orange") into one of the plain-generation pages that also covers an edit-capable gate variant (Flux 2 Max/Pro Edit, Nano Banana Edit, GPT Image 2 Image to Image, Wan 2.6 Image to Image) without attaching a reference image. With no image_urls supplied, you get a brand-new unrelated image, not an edit of anything.

Frequently asked questions

Related prompting guides