Neta Lumina: Anime-Style Image Generation in ComfyUI

Neta Lumina is a highly expressive anime-style model fine-tuned on the original Lumina diffusion checkpoint. This fine-tune was created to enhance the rendering of sensual, fantasy-rich, and stylized character imagery, giving artists and prompt engineers a new playground for narrative-rich generations.

This guide is designed for creators using ComfyUI to get the most out of Neta Lumina's abilities — from mastering prompt syntax to controlling character expression, pose, and styling. Whether you're crafting glowing cyberpunk girls, mysterious shrine maidens, or pixel-art idol scenes, this tutorial provides you with clear strategies and examples.

Neta Lumina is a high-quality anime-style image generation model developed by Neta.art Lab, fine-tuned on the open-source Lumina-Image-2.0 from the Alpha-VLLM team at Shanghai AI Laboratory.

It has been further trained using a vast collection of high-quality anime artworks and multilingual tag data. The result is a model with excellent comprehension, creative expression, and prompt adherence — making it especially strong for use cases like illustration, storyboarding, posters, character concept design, and more.

Key Highlights:

• Powered by Gemma text encoder for deep understanding of prompts.

• Multilingual input support: optimized for English, Chinese, and Japanese.

• Versatile visual coverage: from popular aesthetics to niche styles (including Guofeng, Furry, and Pet themes).

• Still supports Danbooru-style tags, but also works great with natural language.

Neta Lumina excels in bringing emotional depth, detailed outfits, and dynamic anime compositions to life — especially when used in tools like ComfyUI.

Before diving into Neta Lumina’s expressive capabilities in ComfyUI, you’ll need to set up your environment correctly. Whether you're running locally or on a cloud GPU platform like RunPod, this section will walk you through everything required to get started.

1. Install ComfyUI

Neta Lumina runs inside ComfyUI, a flexible and powerful node-based interface for AI image generation.

• Local Installation
  Ideal if you have a powerful PC or workstation (12GB+ VRAM recommended).
  👉 Follow our Local ComfyUI Setup Guide.

• RunPod Deployment
  Best for users who prefer cloud GPU access with minimal setup.
  👉 How to run ComfyUI on RunPod.

2. Download the Required Files

To run Neta Lumina, you'll need to manually download these components:

Model Checkpoint (UNet)

• 🔗 neta-lumina-beta-aes-experimental-0624.safetensors

• 📁 Save to: ComfyUI/models/unet/

Text Encoder (Gemma-2B)

Neta Lumina uses the Gemma-2B text encoder for deep prompt comprehension.

• 🔗 gemma_2_2b_fp16.safetensors

• 📁 Save to: ComfyUI/models/text_encoders/

Variational Autoencoder (VAE)

Neta Lumina requires a specific VAE (same as Flux VAE) for decoding latent representations into final images.

• 🔗 ae.safetensors

• 📁 Save to: ComfyUI/models/vae/

3. Verify Your Setup

After downloading and placing everything in the correct folders:

ts
1📂 ComfyUI/
2├── 📂 models/
3│   ├── 📂 unet/
4│   │   └── neta-lumina-beta-aes-experimental-0624.safetensors
5│   ├── 📂 text_encoders/
6│   │   └── gemma_2_2b_fp16.safetensors
7│   └── 📂 vae/
8│       └── ae.safetensors

With ComfyUI installed and the Neta Lumina files in place, you're ready to begin loading the workflow and prompting like a pro.

Now that your environment and model files are in place, it’s time to load and configure the Neta Lumina workflow in ComfyUI. This setup ensures that all components work together correctly for optimal image quality, stability, and performance.

1. Load the Neta Lumina Workflow

Start by loading the Neta Lumina workflow:

👉 Download the official Neta Lumina Base workflow JSON file and drag it into your ComfyUI canvas.

This template includes all required nodes pre-arranged for easy customization and testing.

2. Configure the Nodes

Double-check each node in the workflow to ensure it references the correct model files and optimal generation settings:

🎉 Once configured, you're ready to prompt with the full power of Neta Lumina!

With your workflow now fully configured, it’s time to put Neta Lumina to the test using a visually rich and highly expressive prompt. To start, we’ll simply use the base prompt included in the workflow, giving you a quick feel for how well the model renders anime-style characters, intricate details, and emotional tone.

💡 For this test, we changed the canvas dimensions to a 16:9 ratio (1280×720) instead of the recommended square or portrait resolutions — and it works great! This render took approximately 12 seconds on an RTX 4090 (24GB VRAM). The first generation may take slightly longer as models initialize.

✨ Positive Prompt:

You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> (masterpiece, best quality:1.2), 1girl, twin tails, multicolored hair, purple eyes, happy, freckles, upper body, from below, white silk dress, smiling, cleavage, large breasts, front view, looking at viewer, cherry blossom, natural light, sitting on beach, ninja stars flying in background, psychedelic art

✨ Negative Prompt:

You are an assistant designed to generate images based on textual prompts. <Prompt Start> (worst quality:2), (low quality:2), (normal quality:2), blurry, jpeg artifacts, watermark, signature, extra limbs, mutated hands, bad anatomy, deformed, disfigured, poorly drawn face, bad proportions, missing fingers, fused fingers, long neck, long body, duplicate, morbid, mutilated, ugly, disgusting, (bad hands:1.5), (bad feet), cross-eyed, low-res, overexposed, underexposed, (monochrome:1.2), (grayscale:1.2), nsfw watermark, censored, out of frame, text

As you can see, the prompts in this workflow use a slightly different structure than normal. Each prompt is prepended with a system-style instruction like:

ts
1
2You are an assistant designed to generate images based on textual prompts. <Prompt Start>

✅ This structure is applied to both positive and negative prompts — and is highly recommended to boost image quality and comprehension.

From First Render to Prompt Mastery

With your first generation complete, you’ve now seen what Neta Lumina can do using the base prompt — a solid test of color dynamics, stylistic fidelity, and subject clarity. To get the most out of it, it’s time to go deeper.

Like many anime-focused models, Neta Lumina was trained on a dataset built from Danbooru-style tags. These are short, descriptive keywords sourced from the anime art platform Danbooru, where images are labeled with detailed attributes like characters, expressions, outfits, camera angles, and environments. Think of these tags as a shared visual language between you and the model.

👉 Next, we’ll explore how to build powerful prompts using these tags, how to weight them for precision, and how to mix aesthetic and technical details to guide your generations exactly where you want them.

Danbooru is a massive anime imageboard where each image is tagged with ultra-specific descriptors — characters, hairstyles, outfits, lighting, camera angles, and more. These tags form the training data behind models like Neta Lumina, and learning to use them properly gives you much greater control over the results.

Unlike natural sentences, prompts here are made of compact, descriptive tags stacked together. Think of it like building an illustration layer by layer.

Use this flexible structure as your guide:

• 🧭 System Instruction (Strongly Recommended)
  A guiding sentence at the top of your prompt that tells the model how to behave.
  Example:

  You are an assistant designed to generate anime images based on textual prompts. <Prompt Start>

• 👤 Character Identity
  Who’s in the image? Choose the number, type, or group.
  Example: 1girl, solo, 2girls, group

• 🎨 Art Style & Quality Tags
  Steer the overall finish and style of the image.
  Example: masterpiece, (best quality:1.2), anime screencap

• 💇 Appearance Details
  Hair color, eye style, body traits, or signature features.
  Example: twin tails, purple eyes, freckles, slim waist

• 👗 Clothing & Accessories
  Pick the outfit, fabric, or extras like hats, glasses, etc.
  Example: white silk dress, school uniform, ribbons, cleavage

• 😊 Expressions & Actions
  What’s the mood or what are they doing?
  Example: smiling, happy, crying, holding umbrella

• 🎥 Camera Angle & Framing
  Perspective and how the viewer sees the subject.
  Example: from below, front view, upper body, close-up

• ✨ FX & Microdetails
  Stylistic touches, magic, particles, glows, or weather.
  Example: natural light, glowing eyes, ninja stars flying

• 🌸 Background & Setting
  Where the scene takes place and what atmosphere it has.
  Example: sitting on beach, cherry blossom, psychedelic art, foggy forest

💡 Tip: You don’t need to use every tag type. Many great prompts use just 5–7 categories, depending on your creative goal.

While tag-based prompts are highly effective on their own, you can also start with a tag skeleton and then enrich it using natural language with tools like ChatGPT-4o. This hybrid approach lets you combine precision with expressive detail in the same prompt.

Now that you know how the tag structure works, it’s time to see it in action. In the next section, we’ll walk through real examples built using this approach — showing how different tag combinations, styles, and categories affect the final image. You’ll see side-by-side outputs that highlight how prompts shape everything from character design to lighting, composition, and emotion.

Now that you understand the structure of Danbooru-style prompts, let’s dive into some real-world examples. These prompts use concise, stacked tags that layer an illustration step-by-step, giving you a clear sense of how to combine character details, art styles, expressions, and more to achieve varied effects. Below each prompt, you’ll find a breakdown showing which tags correspond to the different categories — a handy way to see the building blocks in action.

✨ Example 1

You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> 1girl, illusion art, white shirt, black pantyhose, pleated skirt, cleavage, outdoors, front view, upper body, sparkling eyes, sparkle background, wings, twin braids, red eyes, blue hair, fox tail, underboob, big boobs, wet skirt, breasts, huge ass, (best quality:1.1)

✨ Example 2

You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> 1girl, vaporwave aesthetic, pastel colors, oversized hoodie, short pink hair, neon sunglasses, relaxed pose, city rooftop, sunset sky, (best quality:1.1)

✨ Example 3

You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> 1girl, dark fantasy, black armor, glowing red eyes, long silver hair, sword drawn, battle-ready, stormy sky, dramatic lighting, full body, (best quality:1.1)

✨ Example 4

You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> (masterpiece, best quality), 1boy, flame sword, cyber armor with exposed wires, glitchy neon rain, city burning behind, screaming, cinematic motion blur, stylized anime lighting

✨ Example 5

You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> (masterpiece, best quality), 1girl, reclining on neon-lit pixel bed, silk robe slipping off shoulder, garters, neon underglow, purple-dyed hair, finger to lips, close-up, ambient room particles, cyberpunk bedroom

✨ Example 6

You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> (masterpiece, best quality), 1girl, sensual pixel art illustration, backless dress, lying on velvet carpet, low angle, winking, candlelight glimmer, dragon tattoo on thigh, rose petals floating in frame

✨ Example 7

You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> (masterpiece, best quality), pixel art, close-up of a cybernetic griffin’s eye, mechanical iris with rotating gears, sharp chrome feathers, neon orange light accents, rain dripping on metallic surface, dystopian cityscape blurred behind, high-contrast sharp shadows

✨ Example 8

You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> (masterpiece, best quality), 1girl, braided pink ponytail, mid-jump dynamic pose, Game Boy color palette, high contrast shading, abstract glitch background, arms wide open, pixelated heart UI elements floating around

✨ Example 9

You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> (masterpiece, best quality:1.2), 1girl, seductive anime art, piercing green eyes, sleek ponytail, full cleavage, confident smirk, front view, classroom bathed in golden hour light, pleated skirt with subtle rips, fitted school uniform blouse, hands firmly on hips, busty, (wide hips:1.45), small waist, voluptuous figure

✨ Example 10

You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> (masterpiece, best quality), 1girl, cowboy hat, bandana covering lower face, worn leather jacket, revolver holster, dusty desert sunset background, sharp black line art, high contrast shadows, minimal shading, sepia tones, gritty sketch style, dynamic pose, confident smirk, wind blowing hair, western manga vibe

✨ Example 11

You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> (masterpiece, best quality), 1girl, sitting under alien tree with translucent leaves, watercolor art style, glowing blue moonlight, wide reflective eyes, soft brush textures, atmosphere of wonder, floating spores around

✨ Example 12

You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> (masterpiece, best quality), 1girl, twin-blade spin attack, front view close-up, intense focus, surrounded by shadow creatures with glowing red eyes in the background, ancient temple ruins glowing with green glyphs, ground cracking underfoot, dynamic motion blur, mid-motion capture frame, dramatic lighting

✨ Example 13

You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> (masterpiece, best quality), 1girl, close-up face, glitch effect splitting her features, neon pink and blue hues, intense expression, cyberpunk city blurred behind

✨ Example 14

You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> (masterpiece, best quality), 1girl, dark horror vibe, demonic wings with glowing veins, spiked tail, pointed cat ears, shadowy smoky background, piercing yellow eyes, ripped dark outfit, ominous atmosphere, chilling presence

✨ Example 15

You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> (masterpiece, best quality), 1girl, black and white art, sensual loose hair flowing over shoulders, large breasts, wide hips, reclining by an indoor hot spring, wet silk robe clinging to her curves, smoky steam swirling in the background, sultry gaze, intricate fine line work, moody deep shadows, leaning forward with a sensual smile, sexy twin tails hairstyle, emphasizing cleavage and inviting pose

✨ Example 16

You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> (masterpiece, best quality), 2girls, both with horns, tails, and glistening dark skin, wearing tiny strappy lingerie that barely covers their curves, lounging confidently on a throne of bones, the other leaning forward to seduce a chained viewer from the edge of the frame, firelight flickering across their glossy skin, glowing eyes, wicked smirks dripping with lust and danger, seductive demon queen vibe, shadows dancing around them

✨ Example 17

You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> (masterpiece, best quality), 1boy, muscular, 30 years old, spiky grey hair, blue eyes, black ninja tactical armor, red headband with fluttering tails, black face mask, intense expression, ninja stars flying toward viewer, forward motion pose, low angle perspective, futuristic battlefield with alien mechs, pixel art, high contrast lighting, anime style

With a variety of prompt styles and concepts at your fingertips, the next challenge is crafting detailed and accurate prompts quickly. To make this easier, we’ll now explore how to integrate Danbooru tag selection directly into ComfyUI, so you can effortlessly find and insert tags as you type.

Now that you have a solid collection of curated prompts and styles to inspire your anime image generation, the next step is to streamline how you craft those prompts. Typing out complex tags manually can be time-consuming and prone to errors, especially with the extensive and nuanced Danbooru tag system. To solve this, we’ll integrate the ComfyUI Custom Scripts Github Repo into ComfyUI that allows you to conveniently select Danbooru tags from a dropdown menu as you type, dramatically speeding up your workflow and improving prompt accuracy.

How to Enable Danbooru Tag Autocomplete in ComfyUI

Follow these steps to enable autocomplete suggestions for Danbooru tags directly inside ComfyUI’s prompt input field.

Step 1: Install ComfyUI-Custom-Scripts

1. In ComfyUI, click on “Manager” at the top right of the interface.

2. Then select “Custom Nodes Manager”.

3. Use the search bar to look for ComfyUI-Custom-Scripts.

4. Click Install, select the latest version if prompted, and confirm.

   

5. After installation, follow the popup instructions at the bottom-left of the screen:

   • Click “Restart Custom Nodes”

   • Then Reboot ComfyUI to apply changes.

Step 2: Load the Danbooru Tag List

1. Back in the main ComfyUI interface, go to Settings ⚙️ (bottom-left corner).

2. Use the search bar to find “autocomplete”.

3. Next Click on “Manage Custom Words”.

   

4. In the “Load Custom List” field, paste this link:

   ts
   1https://cdn.nextdiffusion.ai/danbooru/tags.txt

5. Click “Load” next to the input field and wait for the tag list to finish importing (this might take a few seconds).

   

6. Once loaded, click “Save” to finalize the configuration.

Step 3: Enjoy Tag Suggestions While Typing

Now, when you type inside any prompt input field in ComfyUI, you’ll see a dropdown with live autocomplete suggestions based on the loaded Danbooru tags.

This dramatically speeds up prompt creation and helps you stay accurate with complex or lesser-known tag names.

With Danbooru tag autocomplete now integrated, crafting high-quality prompts becomes much faster and more reliable. But beyond general tags, Neta Lumina supports special structured tags that unlock even more control — especially for anchoring specific artist styles and character identities. Let’s explore how to use them properly!

In addition to standard Danbooru-style tags, Neta Lumina supports structured tags that help you lock in a specific character identity or artist style more effectively. These tags are powerful tools to bring recognizable personalities or signature aesthetics into your AI-generated art.

What’s the difference?

These tags are not just cosmetic—they directly influence the model’s behavior during generation, especially when placed early in the prompt or near the system prompt.

Real-World Usage Example

Example Image:

You are an assistant designed to generate images based on textual prompts. <Prompt Start> Characters: #hatsune_miku. Drawn by @ebifurya. 1girl, running forward dynamically, wind tossing hair and ribbons, playful yet sensual smile, determined look, detailed lighting, masterpiece, best quality, chaotic post-apocalyptic battlefield with bright explosions and smoke plumes

💡 This tells the model:

• Who to draw → Hatsune Miku

• How to draw → In Ebifurya's signature artistic style

To get a really distinct style from @ebifurya—which is known for its smooth, painterly, anime-style digital art with soft shading and detailed character rendering—let’s try a contrasting sticker art style by @bkub to see how dramatically the artist tag influences the output.

Example Image:

You are an assistant designed to generate images based on textual prompts. <Prompt Start> Characters: #hatsune_miku. Drawn by @bkub. 1girl, running forward dynamically, wind tossing hair and ribbons, playful yet sensual smile, determined look, detailed lighting, masterpiece, best quality, chaotic post-apocalyptic battlefield with bright explosions and smoke plumes

💡 This tells the model:

• Who to draw → Hatsune Miku

• How to draw → In Bkub’s distinctive sticker art style, featuring bold outlines, flat vibrant colors, and a playful, cute aesthetic reminiscent of modern digital sticker

Tag Formatting Rules (Important)

To ensure your structured tags are parsed and interpreted correctly by the model and interface -especially in tools like ComfyUI, follow these rules:

• Always keep underscores (_) in character and artists names
  Many character and artist names in Danbooru-style datasets use underscores. Removing or replacing them (like with spaces or hyphens) can break the reference.
  ✅ #hatsune_miku
  ❌ #Hatsune Miku or #hatsune-miku

• Only escape parentheses if they are part of the tag
  Some artist names (especially niche ones) include parentheses as part of their official tag. In those cases, you must escape them with backslashes so they aren’t misinterpreted by the parser.

  When to escape:
  ✅ @quan_\(kurisu_tina\)
  This is required if the artist tag is literally: quan_(kurisu_tina)

  When not to escape:
  ❌ @ebifurya → doesn’t need escaping because it has no parentheses.
  ❌ #hatsune_miku → also fine as-is.

• Use parentheses for disambiguation (no escaping needed)
  Tags like #rem_(re_zero) are common for characters with shared names. The parentheses here are part of the tag — do not escape them unless you're using the @ symbol and it's needed by your interface.

With a clear understanding of how structured tags like character and artist identifiers shape your AI-generated images, you’re ready to explore the next step in perfecting your creations: enhancing image clarity and detail with the Neta Lumina Upscale ComfyUI Workflow.

Upscaling is an excellent way to boost the quality of your anime images by sharpening details and increasing resolution without losing stylistic fidelity. To get started with the Neta Lumina Upscale Workflow in ComfyUI, follow these steps carefully.

1. Download the Required Files

To run the Neta Lumina Upscale Workflow, you’ll need to download the primary model checkpoint used for anime image upscaling.

Upscale Model Checkpoint:

• 🔗 4x-AnimeSharp.pth

• 📁 Save to: ComfyUI/models/upscale_models/

2. Verify Your Setup

Make sure the downloaded model file is in the correct folder to avoid errors during workflow execution. Your folder structure should look like this:

ts
1📂 ComfyUI/
2├── 📂 models/
3│   └── 📂 upscale_models/
4│       └── 4x-AnimeSharp.pth

With the model in place, load the Neta Lumina Upscale workflow into ComfyUI to start upscaling your images.

3. Load the Neta Lumina Upscale Workflow

👉 Download the provided Neta Lumina Upscale workflow JSON file and drag it into your ComfyUI canvas.

This workflow includes all the necessary nodes, pre-arranged for quick customization and testing.

4. Install Missing Node: Ultimate SD Upscale

If a missing node warning appears when loading the Neta Lumina Upscale workflow, it’s because the UltimateSDUpscale node isn’t installed yet.

To fix this:

1. Click Manager in the top-right corner of ComfyUI.

2. Select Install Missing Custom Nodes.

3. You should see ComfyUI-UltimateSDUpscale listed — simply install the latest version.

   

4. After installation, click Restart and Reboot (bottom-left) when prompted.

Once restarted, reload the workflow and the node will be properly loaded.

5. Run ComfyUI Neta Lumina Ultimate SD Upscale Workflow

You're now ready to use the Neta Lumina Upscale Workflow. Simply enter a prompt, set your desired image dimensions (width/height), and adjust key settings inside the UltimateSDUpscale node for best results.

### Key Settings Overview

Core Settings:

• upscale_by: 2.0 – Multiplies the resolution (2x recommended).

• steps: 30 – Denoising steps; higher = better quality.

• cfg: 8.0 – Prompt adherence strength.

• denoise: 0.10 – Lower values keeps the original image more intact.

Tiling:

• tile_width / tile_height: 512 – Smaller tiles = lower VRAM use.

• tile_padding: 32 – Adds overlap to reduce seams.

• mask_blur: 8 – Softens edges between tiles.

Seam Fixing (Optional):

• seam_fix_mode: None – Can be toggled on if needed.

• seam_fix_denoise: 1.0 – Strength of seam blending.

💡 Tip: To maintain visual consistency during upscaling, match the sampler_name and scheduler with your original KSampler settings. For this workflow, use sampler_name: res_multistep and scheduler: linear_quadratic.

That’s it! Tweak the parameters to match your GPU and output needs, then run the workflow to upscale your images with pro-level clarity.

6. See the Difference – Neta Lumina Upscale Example

We’ve covered everything you need to get the most out of the amazing Neta Lumina anime model—from setup in ComfyUI to crafting powerful prompts using Danbooru-style tags and special artist and character tags. You’ve also seen how the custom scripts for easy Danbooru tag autocomplete make prompt building a breeze, plus the upscale workflow to enhance image clarity and bring out even finer details. With all these tools, you can create stunning, vibrant anime-inspired images with impressive style and precision.

✨ Future Improvements

• 🎯 Better anatomy (hands & feet) accuracy for more natural and dynamic character poses

• 📝 Improved interpretation of text within images, such as signs or written elements

Neta Lumina already offers a fantastic foundation for creative expression, and as the model continues to evolve, your art will become sharper, more expressive, and even more fun to make.

Get ready to bring your anime creations to life like never before!