What is Lumina-Image 2.0?

Lumina-Image 2.0 is a 2.6B parameter image generation model that combines Unified Next-DiT and Flux-based VAE.

While adopting the Gemma 2B text encoder, the model body is considerably smaller than SD3 and FLUX Pro, and like AuraFlow, it is designed aiming for the "relatively lightweight and easy-to-use base model" category. It is also characterized by high prompt adherence for its size, and attracted attention as one of the candidates for the next-generation base model.

However, since it uses Gemma 2B (2B parameters) as a text encoder, it should be noted that the VRAM usage for the text encoder is slightly larger compared to SD1.5 etc.

Model Download

📂ComfyUI/
└──📂models/
    ├── 📂diffusion_models/
    │   └── lumina_2_model_bf16.safetensors
    ├── 📂text_encoders/
    │   └── gemma_2_2b_fp16.safetensors
    └── 📂vae/
        └── ae.safetensors

text2image

Lumina-Image-2.0.json

Neta Lumina

Neta-Lumina is a fine-tuned model for anime based on Lumina-Image 2.0.

Like an anime model, it also supports Danbooru tags and is characterized by accepting multi-language prompts such as Chinese, English, and Japanese.

Model Download

📂ComfyUI/
└──📂models/
    └── 📂diffusion_models/
         └── neta-lumina-v1.0.safetensors

text2image

neta-lumina-v1.0.json
  • Follow the official settings for the sampler, use res_multistep / linear_quadratic.

Prompts are a bit unique, and you need to write a system prompt before the text you actually want to generate.

You are an assistant designed to generate anime images based on textual prompts. <Prompt Start>
1girl, portrait, ...

Please refer to the official Prompt Book for details.

NetaYume Lumina

There is also a model called NetaYume Lumina, which is further fine-tuned based on Neta Lumina.

I will introduce this as well.

Model Download

📂ComfyUI/
└──📂models/
    └── 📂diffusion_models/
         └── NetaYumev4_unet.safetensors

text2image

NetaYumev4.json

NewBie image Exp0.1

NewBie-image (Exp0.1) is an anime-focused T2I model with a unique NewBie architecture designed based on Next-DiT, incorporating insights from Lumina architecture research. It uses a more powerful text encoder and enables more detailed control with XML-formatted prompts (structured tags).

This model is only 20% trained. The workflow may change with future updates.

Model Download

📂ComfyUI/
└──📂models/
    ├── 📂diffusion_models/
    │   └── NewBie-Image-Exp0.1-bf16.safetensors
    ├── 📂text_encoders/
    │   ├── gemma_3_4b_it_bf16.safetensors
    │   └── jina_clip_v2_bf16.safetensors
    └── 📂vae/
        └── ae.safetensors

text2image

NewBie_image_Exp0.1.json

Prompts in XML format (structured with tags) are recommended.

<general_tags>
  <style>
    anime_style, key_visual, official_art, illustration,
    refined_lineart, clean_lineart, high_contrast
  </style>
  <background>
    underwater, deep_blue_water, water_surface, waterline,
    caustics, light_rays, reflections
  </background>
</general_tags>

However, you can generate images without problems even if you write in natural language, so please feel free to try it first.

Please refer to the official prompt guide for details.