Prompting
The header image above was created with FooocusPlus 1.0.9-13, utilising Flux1.Dev with the following prompt along with the TC Coloured Pencil Full Art style. a grey traditional styled alien dressed as a garden gnome holding a piece of chalk, standing in front of a large chalk board on a wall, covered in scribbled random chalk written gibberish words as if being punished in school
The following is a fairly comprehensive how-to guide on effective prompting for major SDXL-based models (including base SDXL, Juggernaut XL, Pony Diffusion, and other SDXL variants) versus Flux series models (Flux.1 dev/schnell/pro variants, including tuned ones like Flux.1 Krea dev), with a focus on usage in FooocusPlus, a fork that supports Flux. This guide highlights key differences in model architecture, prompt understanding, and best practices for success. Flux models represent a major leap in prompt adherence, natural language handling, and output quality compared to SDXL-era models, but they require adjusted techniques—especially since negative prompting is not supported (and often ineffective or ignored) in Flux.
Core Differences Between SDXL-Family and Flux-Family Models
| Base Model | SDXL-Family: SDXL, Elsewhere (the Default), Juggernaut, Pony, etc. | Flux-Family: Flux.1, Krea dev, Schnell, etc. |
| Architecture | Based on Stable Diffusion XL (latent diffusion, CLIP + OpenCLIP encoders) | Flow-matching + advanced transformers (T5 + CLIP), much larger context, better coherence |
| Prompt Style | Works best with comma-separated tags, weights (e.g., (keyword:1.2)), artist names, quality boosters | Excels at natural language sentences, descriptive paragraphs; tags/commas less necessary or even counterproductive |
| Prompt Length | Medium-long (75–150 *tokens); too long dilutes adherence | Handles long, detailed prompts extremely well (200+ *tokens often better) |
| Negative Prompt | Very useful; common lists include blurry, deformed, ugly, low quality, extra limbs. Negative prompting is also available in Fooocus Enhance for realism and Fooocus Semi Realistic. | Not supported — Flux ignores or struggles with negation; use positive rephrasing instead |
| Prompt Adherence | Good but often requires heavy weighting, styles, and fixes for complex scenes/anatomy | Excellent — follows complex instructions, poses, compositions, text rendering far better |
| Strengths | Vast ecosystem (LoRAs, fine-tunes), fast generation, great for stylised/anime (Pony) or realism (Juggernaut) | Superior realism, anatomy, hands, text in image, lighting, fewer artifacts; more opinionated/cinematic |
| Weaknesses | Can struggle with exact poses, text, multi-subject scenes without tweak | Slower generation (especially dev), higher VRAM needs; less flexible with extreme stylisation without tuning |
| FooocusPlus Notes | Native support; auto-expands prompts, styles work well, negative prompt effective | The FooocusPlus negative prompt field is often ignored for Flux |
*A beginners note on 'Tokens' - Instead of reading text word-for-word, AI breaks down input (prompts) and output into smaller chunks — such as characters, partial words, or full words — known as tokens.
Best Practices for SDXL-Based Models (Including Juggernaut XL, Pony, Elsewhere etc.) These models thrive on structured, weighted prompts. In Fooocus, use the prompt box + styles + quality sliders
Prompt Structure — Start with subject -- details -- style -- quality.
Example base: 1girl, solo, beautiful woman with long flowing red hair, blue eyes, wearing elegant Victorian dress, standing in misty forest at golden hour, highly detailed, masterpiece. Add weights for emphasis: (masterpiece:1.3), (best quality:1.2), intricate details
For Juggernaut XL (photo-realism-focused): Use trigger words like skin textures, detailed skin, realistic, photo-realistic, sharp focus. Keep first sentence clear and important. Avoid over 75 tokens for best adherence. Include lighting (soft natural light, dramatic shadows) and perspective (close-up portrait).
For Pony Diffusion (anime/stylised/furry) Use score tags at start/end like score_9, score_8_up for quality. Natural language or tags work; longer detailed prompts beat short ones. Emphasise anatomy/actions with verbs (sitting, dynamic pose) score_9 etc is included in the Fooocus Pony and Pony Real styles - and in the Pony presets.
Negative Prompt — Essential. Common effective list: blurry, lowres, deformed, ugly, mutated hands, extra limbs, poorly drawn face, bad anatomy, watermark, text, signature
Tips for Success:
Use Fooocus styles/presets (e.g., Fooocus V2, Cinematic) to auto-boost. Weight important elements: (red hair:1.4)
Experiment with samplers (Euler a, DPM++ 2M Karras) and 20–40 steps.
For realism (Juggernaut): Add raw photo, 8k or artist references sparingly.
For anime/Pony: Include anime style, detailed background and avoid realism boosters.
Best Practices for Flux Series Models (Flux.1 dev/schnell/pro, Krea dev in FooocusPlus) Flux prefers human-like descriptions. Negative prompts are not available at this point so do nothing (or harm) — avoid them entirely in FooocusPlus for Flux. Instead, positively describe what you want (or rephrase to exclude via inclusion of opposites).
Prompt Structure — Write naturally, like describing a scene to a director.
Example: A cinematic portrait of a young woman with long wavy auburn hair and emerald green eyes, wearing a flowing white dress, standing confidently on a cliff-side at sunset with dramatic golden lighting and ocean waves crashing below, sharp focus, highly detailed, professional photography style
Be explicit about placement/composition: foreground: woman facing camera, background: misty mountains, object A in front of object B
Specify lighting, mood, and angles: dramatic side lighting, low angle shot, serene atmosphere
For text in image: Flux excels — just include sign reading 'Welcome' in bold red letters
Negative Prompt — Leave blank or ignore the field in FooocusPlus. Rephrase negatives positively: Instead of no blur: sharp focus, crystal clear details Instead of no extra limbs: perfect anatomy, two arms, two legs Avoid white background — it often blurs; say detailed colourful background instead.
Tips for Success: Use natural sentences over tags/commas — Flux understands context better. Longer, descriptive prompts = better results (no token limit worry like SDXL). Flux is opinionated (especially Krea dev) — it adds tasteful interpretation; embrace it for cinematic outputs. Settings in Fooocus forks: 15–30 steps typical (schnell faster/lower steps); CFG 3.5–7 (lower for creativity). For variants like Z-Image-Turbo: Similar to base Flux but faster — prioritise concise yet descriptive prompts. Great for realism, complex scenes, hands/text — fewer fixes needed vs SDXL.
Eye-Level: Camera is at the subject's eye level, providing a neutral, standard viewpoint.
High-Angle: Camera looks down on the subject, making them appear vulnerable, weak, or frightened.
Low-Angle: Camera looks up at the subject, making them appear powerful, heroic, or intimidating.
Bird's Eye / Aerial: Directly overhead, providing a top-down, panoramic view often used for context.
Dutch Angle / Tilt: The camera is tilted, creating a disorienting, tense, or psychological effect.
Over the Shoulder (OTS): Positioned behind one person, looking at another, ideal for dialogue.
Point of View (POV): Captures the scene from a character's perspective.
Ground/Knee/Hip Level: Placed very low to the ground, focusing on feet, action, or a dramatic upward view.
Start simple, iterate: Generate, tweak prompt, regenerate with same seed for comparison.
Use Fooocus auto-expansion (if enabled) for SDXL — it adds beauty boosters.
For Flux: Rely on prompt alone more; avoid over-relying on styles/LoRAs (support varies).
Test both families: SDXL for speed/customisation, Flux for superior quality/adherence.
Presets are an ideal way to get started, there is a Wiki page here for your reference and guidance.
Styles are also another way to improve your image output, you can find more information and demonstrations of Styles, e.g. what they do etc here at this Wiki page.
Community resources if you get stuck: Ask here at Pure Fooocus AI or check Civitai for model-specific example prompts.