SD vs Midjourney vs NovelAI

How the same prompt behaves across three major models — comparing how Promgrammer tags respond on SD, Midjourney, and NovelAI.

한국어 English

1. Why the same prompt produces different results

A prompt like masterpiece, 1girl, long silver hair, blue eyes, school uniform produces markedly different images on SDXL, NovelAI, and Midjourney — not because the subject is ambiguous, but because each model was trained differently. Three factors dominate the variance: the text encoder (how words turn into numbers), the training data distribution (what the model has seen), and the prompt parser (how the UI splits and weights your tokens). This guide walks through the six most used models and what to expect from each.

2. Stable Diffusion XL (SDXL base)

Strengths: Photorealism, environmental scenes, wide subject coverage, predictable prompt parsing.
Weaknesses: Anime style requires a fine-tune (Animagine, Pony) — base SDXL anime looks stiff. Weaker on small text and fingers.
Prompt style: Comma-separated tags. Responds to photographic vocabulary (85mm, f/1.8, Kodak Portra) very well.
Negative prompts: Fully supported, commonly used.
Gotcha: masterpiece, best_quality biases toward painting-style output. For photos, drop those tokens.

3. Flux.1-dev / Flux.1-pro

Strengths: Natural-language understanding, in-image text, photo-realistic portraits, prompt adherence.
Weaknesses: Slow, expensive on GPU. Anime style is weak — no widespread fine-tunes yet.
Prompt style: Write full sentences. "A young woman with long silver hair wearing a navy school uniform, standing under cherry blossoms at sunset" outperforms comma-tag style.
Negative prompts: Not supported on Flux.1-dev/schnell. Use positive phrasing ("clear skin" instead of negative "blemishes").
Gotcha: Prompts under 20 words can feel undercooked. Give Flux at least two full sentences.

4. NovelAI v3 / v4

Strengths: Best-in-class anime rendering; strong character consistency; intentional artist tag support.
Weaknesses: Paid service only (no local inference). Photorealism is not a priority.
Prompt style: Heavy on Booru-style tags with underscores or spaces. masterpiece, best quality, 1girl at front is mandatory.
Negative prompts: Supported and crucial. Default UC presets exist.
Gotcha: {tag} syntax increases weight; [tag] decreases. Different from A1111's (tag:1.2).

5. Animagine XL / Counterfeit XL

Strengths: Open-source anime SDXL fine-tune. Free, runs locally, high quality.
Weaknesses: Less prompt flexibility than NovelAI. Some character leakage from Booru tags.
Prompt style: Similar to NovelAI; Booru-tag heavy. masterpiece, best quality, very aesthetic, absurdres is a canonical opening.
Negative prompts: Supported. lowres, bad anatomy, bad hands, text, error, missing fingers covers most cases.
Gotcha: Character trigger words (well-known anime characters) are very strong — even a single token like hatsune miku overrides other features.

6. Pony Diffusion V6

Strengths: Most flexible style range — anime, furry, cartoon. Strong pose rendering.
Weaknesses: Requires the score_9 system instead of masterpiece. Default aesthetic leans rougher than Animagine.
Prompt style: Open with score_9, score_8_up, score_7_up. Then comma tags.
Negative prompts: Supported. score_4, score_3, score_2, score_1, worst quality, low quality common.
Gotcha: Source tag (source_anime, source_cartoon, source_furry, source_pony) is a strong style switch.

7. Midjourney v6 / v7

Strengths: Highest default aesthetic. Excellent compositions with minimal prompting. Coherent faces and hands.
Weaknesses: Less control. Subscription only. Discord UX.
Prompt style: Natural language OR tags — both work. Parameters use --flag syntax: --style raw, --stylize 100, --ar 16:9, --v 6.
Negative prompts: Use --no flag: --no text, watermark.
Gotcha: masterpiece, 8k is ignored. Rely on the --stylize range (0–1000) to control aesthetic intensity.

8. Side-by-side: the same prompt across six models

Consider this Promgrammer-assembled prompt:

1girl, solo, long silver hair, blue eyes, school uniform, cherry blossoms, spring, gentle smile, looking at viewer

SDXL base

Reads as semi-realistic illustration. Facial rendering is soft, closer to concept art than anime. Cherry blossoms render well.

Flux.1-dev

Interprets literally — the result leans photorealistic woman, because school uniform is taken as a real garment, not an anime trope. Rewrite as prose for anime result: "An anime-style illustration of..."

NovelAI v3

Produces canonical anime illustration — the prompt reads exactly as intended. Add masterpiece, best quality to the front.

Animagine XL

Similar to NovelAI, very close match. Slightly more modern rendering, sharper linework.

Pony V6

Expects score_9, score_8_up, source_anime prefix. Without it, the aesthetic is flatter.

Midjourney v6 (--niji 6)

With the Niji model, produces polished anime. Needs fewer tokens; the last few tags (cherry blossoms, spring, gentle smile) are enough.

9. Weighting syntax cheat sheet

Intent	A1111 / SDXL UIs	NovelAI	Midjourney
Emphasize	`(tag:1.2)`	`{tag}`	`tag::2`
De-emphasize	`(tag:0.8)` or `[tag]`	`[tag]`	`tag::0.5` or `--no tag`
Remove	Negative prompt	UC field	`--no tag`
Split regions	`AND`, `BREAK`	Separator not supported	`::` as separator

10. Which model should I choose?

→ Photorealistic portraits / products: SDXL (RealVis, Juggernaut) or Flux.

→ Clean anime illustration: NovelAI if you pay, Animagine XL if you run local.

→ Stylistic variety (anime + cartoon + furry): Pony V6.

→ Polished posters with minimal prompt engineering: Midjourney.

→ Text in images (shop signs, book covers): Flux or Midjourney.

→ Character sheets / pose control: SDXL + ControlNet pipeline.