How to run Lumina Image 2.0 locally.

Which model.

All three share one architecture: a 2B flow transformer from Shanghai AI Laboratory’s Alpha-VLLM team, with Gemma 2 2B as its text encoder and the 16-channel Flux.1 VAE.

• Lumina Image 2.0 (January 2025) is the general base model.
• Neta Lumina (v1.0, July 2025) is Neta.art Lab’s anime fine-tune, trained on more than 13 million anime images. It reads Danbooru tags and plain text in Chinese, English and Japanese.
• NetaYume Lumina is a community fine-tune of Neta Lumina by duongve, now at v4.0 (December 2025). Comfy’s anime template uses its v3.5, and v4 also understands JSON and XML structured prompts.

For anime, start with NetaYume. For anything else, Lumina Image 2.0.

Files you need.

The easy way is one all-in-one checkpoint that goes in checkpoints. Split files use about 2 GB less VRAM.

All-in-one

• NetaYume NetaYume_v4_all_in_one.safetensors ComfyUI/models/checkpoints/ Comfy’s template uses NetaYumev35_pretrained_all_in_one
  10.6 GB Download
• Neta Lumina neta-lumina-v1.0-all-in-one.safetensors ComfyUI/models/checkpoints/
  10.6 GB Download
• Lumina 2.0 lumina_2.safetensors ComfyUI/models/checkpoints/
  10.6 GB Download

Split files (Lumina Image 2.0)

• Model lumina_2_model_bf16.safetensors ComfyUI/models/diffusion_models/ NetaYume and Neta Lumina have their own in Unet/ folders, 5.2 GB each
  5.2 GB Download
• Text encoder gemma_2_2b_fp16.safetensors ComfyUI/models/text_encoders/ the same file for all three
  5.2 GB Download
• VAE ae.safetensors ComfyUI/models/vae/ the Flux.1 VAE
  0.3 GB Download

models/
├── checkpoints/
│   └── NetaYume_v4_all_in_one.safetensors     (all-in-one)
├── diffusion_models/
│   └── lumina_2_model_bf16.safetensors        (or split files)
├── text_encoders/
│   └── gemma_2_2b_fp16.safetensors
└── vae/
    └── ae.safetensors

Smaller files: GGUF

ComfyUI-GGUF loads Lumina 2 models through its Unet Loader. The Gemma 2 encoder stays a safetensors file.

calcuis also has a Neta Lumina Q4_K_M at 2.0 GB. No official fp8 release exists.

What fits your computer.

At 1024 × 1024. These models compute in bf16 only; forcing fp16 gives black images.

• 6 GBTight

  A GGUF model with the 5.2 GB Gemma encoder partly in system RAM.

• 8 GBFits

  Split files. The NetaYume author puts them at about 8 GB, and Neta’s card asks for 8 GB or more.

• 12 GBFits

  The all-in-one checkpoint, about 10 GB in use.

• 16 GBFits

  All-in-one with room for LoRAs and larger sizes.

• RTX 20 and olderSlow

  No bf16 on these cards, so ComfyUI runs in fp32. An RTX 2060 took 14 s per step.

• MacLikely

  An early Apple Silicon crash was fixed in ComfyUI in February 2025. No Mac timings found.

• 8 GB graphics cards
• 12 GB cards
• 16 GB cards
• Mac with Apple Silicon

Set it up.

1. 1

   Update ComfyUI

   Lumina 2 has been native since February 2025. Use an official, current ComfyUI; old or repackaged bundles fail with “Could not detect model type”.

   Terminal, in the ComfyUI folder Copy

   git pull
   pip install -r requirements.txt

2. 2

   Download a checkpoint

   For a first run, one all-in-one file from the list above into checkpoints. Restart ComfyUI.

3. 3

   Open a workflow

   For NetaYume, pick NetaYume Lumina Text to Image in the template browser. For Lumina Image 2.0, drag in the workflow image from ComfyUI’s Lumina 2 example.

4. 4

   Check the shift

   Both workflows put ModelSamplingAuraFlow before the KSampler: shift 6 for Lumina Image 2.0, 4 in the NetaYume template.

5. 5

   Keep the system prompt

   The prompt box starts with a fixed sentence ending in <Prompt Start>. Write your prompt after it and leave the sentence as it is. For Lumina, the CLIP Text Encode (Lumina 2) node can add it for you instead.

6. 6

   Run

   The first run loads Gemma and the model. The next ones are quicker.

The system prompt.

Lumina 2 was trained with a system prompt in front of every caption. It still draws without one, but Comfy’s workflows and the official demo use it. ComfyUI’s plain text encoder sends exactly what you type, so the prefix has to be in the box. These are the ones the Comfy workflows use.

You are an assistant designed to generate superior images with the superior degree of image-text alignment based on textual prompts or user prompts. <Prompt Start> 

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

You are an assistant designed to generate low-quality images based on textual prompts <Prompt Start> 

For NetaYume, put artist tags as @name, mix Danbooru tags with sentences, and see Neta’s prompt book for more. For readable text in the image, NetaYume’s notes suggest a different prefix that tells the model to render quoted text verbatim.

Settings that work.

Lumina Image 2.0

Steps

36Comfy example: 25. Card: 50

CFG

4

Sampler

res_multistep

Scheduler

simple

Shift

6ModelSamplingAuraFlow

Size

1024 × 1024

Negative

yes

System prompt

superior

Comfy’s example workflow runs 25 steps, and its own note says the official way to sample is shift 6 with 36 steps. Alpha-VLLM’s diffusers card uses 50. To mimic the official CFG truncation and normalisation, ComfyUI has a RenormCFG node.

Neta Lumina and NetaYume

Steps

30NetaYume card: 40 to 50

CFG

4 to 5.5

Sampler

res_multistep

Scheduler

linear_quadratictemplate: simple

Shift

4ModelSamplingAuraFlow

Size

1024 × 1024or 1024 × 1536

Negative

yes, with prefix

System prompt

anime

The Neta and NetaYume cards recommend linear_quadratic; Comfy’s NetaYume template uses simple, which also works. euler_ancestral is the cards’ other sampler. There’s no working few-step version of these models; a community CFG-distill LoRA runs at 20 steps and CFG 1.

When it goes wrong.

ERROR: Could not detect model type of: …lumina_2.safetensors

ComfyUI is too old, or a repackaged bundle. Use a current official ComfyUI and Comfy’s example workflow.

repeat(): Not supported for complex yet on a Mac

Fixed in ComfyUI in February 2025. Update.

Black images with --fp16-unet

Lumina 2 overflows in fp16. Remove the flag and let it run in bf16.

NetaYume draws odd styles or garbage

SageAttention doesn’t agree with it. Start ComfyUI without SageAttention.

ComfyUI crashes in the KSampler with a very long prompt

A known issue with extremely long prompts. Shorten it.

Results don’t match the official demo

Use shift 6 at 36 steps and keep the system prompt. The noise also differs between CPU and GPU.

Questions.

Sources: Lumina Image 2.0 model card, Neta Lumina model card, NetaYume model card, ComfyUI Lumina 2 example, demo comparison thread, fp16 and RTX 2060 thread, NetaYume sampling thread, ComfyUI issue #6700.