0xRavenBlack/image-prompting-methods-2026

Image Prompting Methods 2026

An instruction-tuning (SFT) dataset that teaches LLMs how to prompt 2026-generation image models — Nano Banana / Nano Banana Pro (Gemini image), GPT-Image-1, Midjourney v7, FLUX.1 Kontext, Seedream 4, Stable Diffusion 3.5, Ideogram 3, Recraft V3 and Qwen-Image-Edit.

Every row is an Alpaca-style instruction → output pair (with optional input) covering the prompting techniques that define this model generation: natural-language prompt dialects, the 5-layer prompt stack, semantic negative prompts, sequential/conversational generation, structured infographic templates, character-consistency workflows, and model-specific parameter control.

• 2,113 rows in a single train split
• 10 columns: 3 Alpaca fields + 7 metadata fields for filtering
• 10 model scopes (9 models + cross-model)
• English, MIT

Why this dataset

Image-model prompting changed fundamentally between 2024 and 2026: tag-soup keyword piles gave way to caption-style natural language, one-shot lottery prompting became conversational iteration, generic quality-tag negatives became semantic negatives, and every major model developed its own prompt "dialect". General prompt datasets still teach 2023-era habits (masterpiece, 8k, ultra-detailed) that actively hurt on modern models. This dataset encodes current, model-specific practice instead.

Row composition

Slice	Rows	What it teaches
Technique knowledge Q&A	115	Curated per-model technique guides: conversational editing (Nano Banana), golden rules of Kontext, --stylize/--style raw/--oref (Midjourney v7), CFG & semantic negatives (SD 3.5), quoting text (Ideogram 3, GPT-Image-1), reference-driven generation (Seedream 4), style-locking (Recraft V3), keep-unchanged phrasing (Qwen-Image-Edit), cross-model method (5-layer stack, front-loading, sequential generation)
Prompt crafting	760	Task → production-ready prompt in the correct model dialect, each with a "why this works" rationale
Prompt improvement	504	A weak/broken prompt (in input) → rewritten prompt + diagnosis of what was wrong
Edit instruction writing	448	Client goal → precise edit instruction with change-plus-preserve phrasing
Multi-step workflows	150	Step-by-step conversational chains (campaigns, character series, restorations, localization)
Technical control	136	Midjourney parameter usage, SD 3.5 semantic negative prompts, negative constraints for chat-style models

Schema

Column	Type	Description
instruction	string	The task: a question, a brief, or an improvement request
input	string	Optional context (e.g. the weak prompt to improve); empty for most rows
output	string	The target response: technique explanation, complete prompt, or edit instruction
model	string	Model scope: nano-banana, nano-banana-pro, gpt-image-1, midjourney-v7, flux-1-kontext, seedream-4, sd-3.5, ideogram-3, recraft-v3, qwen-image-edit, all
category	string	Task family: prompt-structure, photorealism, image-editing, design-and-branding, typography-and-text, multi-step-workflow, negative-prompting, technical-parameters, art-and-illustration, infographics, character-consistency, text-to-image
difficulty	string	beginner / intermediate / advanced
technique	string	The 2026 prompting method demonstrated (e.g. semantic negative prompts, 5-layer prompt stack)
tags	list[string]	Fine-grained filter tags (model, domain, method)
source	string	Provenance class: official_documentation, community_best_practice, curated_synthesis
language	string	Always en

Model coverage

Model	Rows	Focus areas
GPT-Image-1	279	instruction following, text rendering, infographics, UI mockups
Seedream 4	268	4K output, reference-driven sets, bilingual/CJK text
FLUX.1 Kontext	250	in-context editing, character consistency, edit chaining
Nano Banana (Gemini 2.5 Flash Image)	237	conversational editing, multi-image fusion, scene preservation
Midjourney v7	236	parameters (--ar, --stylize, --style raw, --oref), draft mode
Stable Diffusion 3.5	208	structured prompts, CFG tuning, semantic negatives
Nano Banana Pro (Gemini 3 Pro Image)	67	typography-heavy designs, 4K masters, multi-zone layouts
Recraft V3	218	vector assets, icon systems, brand consistency, mockups
Ideogram 3	168	posters, menus, packaging, wordmarks, long-text layouts
Qwen-Image-Edit	170	text editing (CJK+EN), pose/style transfer, canvas extension
Cross-model	12	the shared 2026 methodology

The 2026 techniques encoded

1. 5-layer prompt stack — subject → style → composition/camera → lighting → technical parameters, ordered by importance.
2. Front-load the subject — first tokens carry the most attention weight.
3. Natural-language dialects — full sentences beat tag-soup on caption-trained encoders; tags survive only as parameter shorthand.
4. Semantic negative prompts — name the specific unwanted concept; drop legacy junk-tag piles.
5. Sequential generation — one constraint per generation/edit, chained with preservation clauses.
6. Structured infographic templates — artifact → layout zones → quoted verbatim strings → type specs.
7. Character consistency — frozen identity sentence + reference anchoring + one changed axis per render.
8. Change-plus-preserve editing — every edit instruction states what changes and what must not move.

Usage

from datasets import load_dataset

ds = load_dataset("parquet", data_files="data/train-00000-of-00001.parquet")["train"]
print(ds[0])

Filter examples:

# only Midjourney v7 advanced rows
ds.filter(lambda r: r["model"] == "midjourney-v7" and r["difficulty"] == "advanced")

# all typography-related rows across models
ds.filter(lambda r: "typography" in r["tags"] or r["category"] == "typography-and-text")

# prompt-improvement pairs (weak prompt given in `input`)
ds.filter(lambda r: r["input"] != "")

# pandas
import pandas as pd, pyarrow.parquet as pq
df = pq.read_table("data/train-00000-of-00001.parquet").to_pandas()
df[df.category == "image-editing"].difficulty.value_counts()

SFT formatting example

def to_chat(ex):
    user = ex["instruction"] + (("\n\n" + ex["input"]) if ex["input"] else "")
    return {"messages": [{"role": "user", "content": user},
                         {"role": "assistant", "content": ex["output"]}]}

Generation methodology

Content was researched from official model documentation and 2026 community best-practice guides, then curated and synthesized into three row families: hand-written technique Q&A; templated-but-varied prompt-crafting/improvement/editing/workflow rows assembled from curated pools of subjects, styles, lighting and copy decks (deterministic generation, deduplicated on the full instruction/input/output triple); and technical parameter/negative-prompt reference rows. Rationale sections are attached to prompt outputs so the model learns why a prompt is shaped that way, not just the surface form.

Limitations

• The dataset teaches prompting knowledge about image models; it does not contain images.
• Prompts are plausible and best-practice-shaped but were not executed against each model API; rendered results will vary by model version.
• Product names and feature sets reflect the 2026 landscape as researched at creation time; verify against current official docs for production use.
• Difficulty labels are heuristic (author-assigned), not empirically calibrated.