Lumina-Image 2.0 is Alpha-VLLM's text-to-image model, released with downloadable weights and inference code. You can obtain the original files from Alpha-VLLM on Hugging Face or use Comfy-Org's documented packaging for ComfyUI. Choose the package that matches your loader before downloading, because these distributions arrange the model components differently. Model card Repository ComfyUI package
What are the key facts about Lumina-Image 2.0?
| Field | Verified detail |
|---|---|
| Developer | Alpha-VLLM, with collaborating institutions listed in its repository. Repository |
| Released | January 25, 2025. Repository |
| Type | Flow-based diffusion transformer for image generation from text. Model card |
| Size or parameters | Model-card prose says 2B; the repository's model table says 2.6B. Neither should be treated as a total for every pipeline component. Model card Repository |
| License and access | Published Lumina weights are labeled Apache-2.0; the Gemma text encoder has its own terms and access conditions. Model card Gemma card |
| Where it runs | Local inference through the project's Python implementation, Diffusers, or documented ComfyUI workflows. Repository ComfyUI example |
What can Lumina-Image 2.0 do?
For someone seeking a download, Lumina's practical advantage is the availability of identified files and documented loaders. Alpha-VLLM publishes the original model, while Comfy-Org explains the file layout for its repackaged version. This gives you a concrete route from model selection to a local workflow. Repository ComfyUI package
The official ComfyUI example supports a combined checkpoint. Download the specified file, place it in the checkpoints directory, and load the workflow embedded in the example image. That is a useful first path when you want to inspect generation through a graphical workflow. ComfyUI example
The original model card also provides a Diffusers example that generates an image and saves it to disk. The download is therefore a model distribution used by an inference application; exporting the resulting picture is a separate step in that application's workflow. Model card
Choose a first-run prompt with simple, visible requirements. For example, ask for a blue bowl beside a folded cloth on a plain table. Your first goal should be confirming that the selected files load and produce a saved image, before testing more demanding compositions.
What are the limitations of Lumina-Image 2.0?
A checkpoint's headline parameter count is not a complete memory requirement. The original repository identifies a Gemma text encoder and a separate VAE in addition to the image model. Record the actual package and precision you use rather than estimating compatibility from the model name alone. Repository
The published descriptions of model size differ: the card says 2B and the repository table says 2.6B. Preserve that distinction when documenting your installation, and identify the source of any size figure you report. Model card Repository
Access conditions can also belong to a dependency. Google's Gemma model page requires accepting its usage terms to access its files. Follow the documented route if your chosen installation downloads Gemma separately; the Lumina model's license does not replace the text encoder's terms. Gemma card
File format and loader must match. The combined ComfyUI checkpoint belongs in a different location from the separately packaged diffusion model, text encoder, and VAE. Renaming a file does not change its contents or make it compatible with an unrelated loader. ComfyUI package
For troubleshooting, identify the failed step first. An authentication failure concerns file access; a missing-file message concerns placement or selection; an allocation error concerns runtime resources. Keep the error and the selected filenames together before changing the workflow.
How do you download Lumina-Image 2.0 for ComfyUI?
Download the combined ComfyUI checkpoint
Start with an installed ComfyUI environment that supports Lumina. ComfyUI's official example points to lumina_2.safetensors and instructs users to place it in ComfyUI/models/checkpoints. Its downloadable example image contains the workflow. ComfyUI example
Hugging Face's CLI can download a selected repository file to a local directory. The following fetches the combined checkpoint from Comfy-Org; it does not start model inference. CLI documentation Checkpoint file
python -m pip install -U huggingface_hub
hf download Comfy-Org/Lumina_Image_2.0_Repackaged \
all_in_one/lumina_2.safetensors \
--local-dir ./lumina-download
Copy lumina-download/all_in_one/lumina_2.safetensors into your ComfyUI checkpoints directory. Then download the workflow image from the official example page and drag it into ComfyUI. Select the downloaded checkpoint, inspect the loaded graph, and generate a test image. Checkpoint file ComfyUI example
Keep the initial graph intact until the first image saves successfully. Once you have that baseline, duplicate the workflow before modifying it. This gives you a known starting point when testing prompts or changing generation settings.
Use separate components only with a matching workflow
Comfy-Org also documents individual model files. The table below follows its published folder structure; it is an alternative arrangement to the combined-checkpoint route. ComfyUI package
| File | Destination within ComfyUI |
|---|---|
lumina_2_model_bf16.safetensors |
models/diffusion_models/ |
gemma_2_2b_fp16.safetensors |
models/text_encoders/ |
ae.safetensors |
models/vae/ |
Use those files when your workflow loads the components separately. Check each loader's selected filename before running. Keep unrelated copies out of your working notes so you can tell which actual files produced the result.
Record the downloaded revision
Hugging Face's download tools support selecting a revision, including a commit identifier. For an installation you need to reproduce, record that revision alongside the filenames and workflow. A repository name by itself does not identify the exact files used in a past experiment. Download documentation
See the sibling Lumina-Image 2.0 overview for architecture and the Diffusers baseline. The ComfyUI complete guide explains the surrounding application concepts.
How does Lumina-Image 2.0 compare with SDXL?
SDXL is a useful local-generation alternative, with official downloadable weights and a documented base-model workflow. Its card describes a base model that can run independently or with a refiner. Lumina's model files and text-encoding components belong to a different pipeline. SDXL card Repository
| Choice | Practical distinction |
|---|---|
| Lumina-Image 2.0 | Use the specific original distribution or documented Comfy-Org package. ComfyUI package |
| SDXL base | Use an SDXL-compatible pipeline and its published model files. SDXL card |
Compare them using the same creative brief, but preserve their separate workflows. Start by checking whether the requested objects and relationships appear correctly, then review materials, lighting, and composition. File compatibility and visual preference are separate decisions.
What should you know before using Lumina-Image 2.0?
Where should I download Lumina-Image 2.0?
Use Alpha-VLLM's official Hugging Face model page for the original distribution. For ComfyUI, its official example links to Comfy-Org's repackaged checkpoint. Model card ComfyUI example
Which ComfyUI folder takes the combined checkpoint?
Place lumina_2.safetensors in ComfyUI/models/checkpoints. The separate-component distribution uses different model folders, so follow the instructions for the package you actually downloaded. ComfyUI example ComfyUI package
How many parameters does Lumina-Image 2.0 have?
Lumina-Image 2.0's official card describes 2B parameters, while its repository table gives 2.6B. The pipeline also uses a separate text encoder and VAE, so retain the source and component scope when reporting either figure. Model card Repository
Can I use the same file in every image generator?
Lumina-Image 2.0 needs a loader that supports its specific distribution. Comfy-Org documents combined and separate-component packages for ComfyUI, while Alpha-VLLM provides its own inference route. ComfyUI package Repository
Sources
- Alpha-VLLM Lumina-Image 2.0 model card
- Alpha-VLLM official repository
- Comfy-Org packaging and folder instructions
- Official ComfyUI Lumina workflow
- Comfy-Org combined checkpoint file
- Hugging Face command-line documentation
- Hugging Face download and revision documentation
- Google Gemma 2 2B model card
- Stability AI SDXL base model card
Top comments (0)