Neta Lumina Prompting: System Tags and Model Comparisons

TL;DR

• Neta Lumina requires a fixed system prefix in every prompt and uses a Gemma 2 (2B) LLM text encoder, giving it native multilingual and natural-language understanding.
• Tag vocabulary uses spaces instead of underscores, escaped parentheses for weighted tags, and special @artist and #character structured tags independent of regular Danbooru-style tags.
• Among anime models, Neta Lumina sits at the high-comprehension end: it outperforms Illustrious XL and Pony Diffusion on complex and multilingual prompts but requires ComfyUI and a newer DiT architecture.

What Prompt Format Does Neta Lumina Require?

Neta Lumina demands a fixed system prefix at the start of every prompt — both positive and negative prompts. The official documentation states: add You are an assistant designed to generate anime images based on textual prompts followed by <Prompt Start>. For negative prompts, the prefix uses "low-quality images" rather than the positive phrasing (neta.art prompt book).

The model also offers alternative system prompts for different input styles: a Danbooru-tags variant that emphasizes tag-to-image alignment, a natural-language variant for detailed prose descriptions, and a structured-summary variant for scene-element lists. Each variant has significantly less training data than the standard version, so results may vary (neta.art prompt book).

How Does the Gemma 2 Text Encoder Change Prompting Compared to CLIP Models?

Unlike Illustrious XL and Pony Diffusion, which use CLIP text encoders inherited from Stable Diffusion XL, Neta Lumina employs Gemma 2 (2B), a full autoregressive language model. This means the model processes prompts as language rather than as embedding tokens, enabling it to handle natural-language descriptions, multilingual input, and complex compositional instructions that CLIP encoders struggle with (neta.art blog).

The tradeoff is that English performs best because Gemma 2 was primarily trained on English data. Chinese and Japanese are effective but may show varying degrees of degradation. The recommended approach is to build prompts primarily in English, supplementing with Chinese and Japanese where needed (neta.art prompt book).

What Are Neta Lumina's Special @artist and #character Tags?

Beyond regular Danbooru-style tags, Neta Lumina provides structured tags for anchoring artist styles and characters with precision. The @artist tag specifies illustration style using phrasing like Drawn by @{artist_name}, Painted by @{artist_name}, or by @{artist_name}. The #character tag specifies characters using formats such as Characters: #{character_name1}, #{character_name2} (neta.art prompt book).

These special tags have distinct formatting rules: underscores must be preserved in original names, and parentheses within tag names must be escaped with backslashes (civitai.com guide 16274).

For another example of alternative workflow approaches, see our guide to flux training tools. Additional context for prompt engineering workflows can be found in the prompt fundamentals guide. For ComfyUI-specific optimizations, see ComfyUI quality of life tips.

How Does Neta Lumina's Tag Vocabulary Differ from Illustrious XL?

Illustrious XL, built on the SDXL UNet architecture and fine-tuned on Danbooru2023, uses standard CLIP-based text encoding and follows conventional Danbooru tag conventions with underscores (HuggingFace model card, Illustrious technical report). Neta Lumina differs in three key ways:

• Underscore-to-space conversion: school_uniform becomes school uniform
• Escaped parentheses: Weighted tags require star \(symbol\) not star_(symbol)
• Structured tags: Special @artist and #character systems not present in Illustrious XL

Additionally, Neta Lumina's Gemma-based encoder can process natural language directly alongside tags, while Illustrious XL relies on CLIP's embedding-based understanding (neta.art blog, Illustrious technical report).

For related workflow context, see ComfyUI quality of life tips.

How Does Neta Lumina's Tag Vocabulary Differ from Pony Diffusion V6?

Pony Diffusion V6 uses the SDXL architecture with standard CLIP text encoding, similar to Illustrious XL. Its model card was gated and inaccessible, so specific differences require inference. Pony Diffusion typically follows standard Danbooru tag conventions with underscores and no special structured tag system (HuggingFace model page — gated, HTTP 401).

Neta Lumina's tagged vocabulary is more explicitly multilingual in design, with the Gemma 2 encoder enabling native Chinese, Japanese, and English processing, while Pony Diffusion and other SDXL models primarily operate in English (neta.art blog).

Where Does Neta Lumina Fit Among Anime Diffusion Models?

Neta Lumina occupies a distinct position. It's built on the Lumina-Image-2.0 DiT framework (2.6B parameters), a flow-based diffusion transformer treating text and image tokens as a unified sequence (Lumina-Image-2.0 paper). This contrasts with Illustrious XL and Pony Diffusion V6, which use the SDXL UNet architecture.

Comparison table:

What Is the Recommended Prompt Structure for Neta Lumina?

The documentation recommends a nine-part structure for optimal results: Character trigger words, Art-Style, Character Appearance, Clothing, Expression & Action, Camera/Perspective, Lighting & Effects, Scene Atmosphere, and Quality Tag. Building a skeleton with tags first and then enriching details in natural language produces strong results (neta.art prompt book).

Common Errors and Fixes

FAQ

How do you write prompts for Neta Lumina? Build a skeleton with tags following the nine-part structure, then optionally enrich with natural language. Always start with the system prefix. Example: You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> 1girl, school uniform, black hair, purple eyes, gentle smile (neta.art prompt book).

What is the system prompt for Neta Lumina? The required prefix is You are an assistant designed to generate anime images based on textual prompts. <Prompt Start>. For negative prompts, the prefix changes to You are an assistant designed to generate low-quality images based on textual prompts (neta.art prompt book).

How does Neta Lumina differ from Illustrious XL in prompting? Neta Lumina uses Gemma 2 (an LLM) as its text encoder instead of CLIP, enabling native multilingual support and natural-language understanding. Its tag vocabulary uses spaces instead of underscores and includes special @artist and #character structured tags. Illustrious XL uses CLIP encoders and conventional Danbooru tag formatting with underscores (neta.art blog, Illustrious XL model card).

Does Neta Lumina understand Chinese and Japanese prompts? Yes. Neta Lumina has native multilingual support for Chinese, English, and Japanese, with English performing best. Chinese and Japanese are effective but may show varying degrees of performance degradation (neta.art blog, neta.art prompt book).

Can I mix Danbooru tags and natural language in Neta Lumina? The model treats tags and natural language as core inputs at the same level, so using one or the other is recommended for stable results. Mixing both is considered an out-of-distribution usage that may produce unpredictable but sometimes surprisingly pleasant effects (civitai.com guide 16274).

Sources

• Neta Lumina Prompt Book — Official prompt writing guide with system prefix, parameter recommendations, nine-part structure, and multilingual examples.
• Neta Lumina Technical Report — Training strategy, dataset details (13M+ images, 4-phase curriculum), aesthetic scoring methods, and labeling pipeline.
• Neta-Lumina HuggingFace Model Card — Apache 2.0 release details, ComfyUI workflow, component downloads, recommended settings.
• Lumina-Image 2.0 paper (arXiv) — 2.6B parameter flow-based DiT with Unified Next-DiT architecture; base framework for Neta Lumina.
• Illustrious XL technical report (arXiv) — SDXL-based anime model trained on Danbooru2023.
• Illustrious XL model card (HuggingFace) — Kohaku XL Beta 5 fine-tune; quality tags, recommended settings.
• Neta Lumina Civitai Article 16274 — Prompt guide with system prefix examples, structured tag syntax, weight guide, and nine-part structure details.