DownloadAndLoadLuminaModel

This is the node that loads the actual image model - Lumina-Next, Alpha-VLLM's flow-matching text-to-image architecture, in the form of a 2B-parameter transformer called NextDiT_2B_GQA_patch2. If DownloadAndLoadGemmaModel is the text half of the pipeline, this is the generation half, and it's the thing your sampler is going to push around.

The download is fully automatic, but it's worth knowing where the files land because that's where people lose them: ComfyUI/models/lumina/<model-name>/consolidated.00-of-01.safetensors. The node uses snapshot_download from HuggingFace, skipping the ema weights and .pth files, so you end up with one consolidated safetensors and a few config bits.

Which model to pick

You get two choices on the model dropdown:

• Alpha-VLLM/Lumina-Next-SFT (default) - the supervised-fine-tuned release, and the one to use for normal generation. It's the "tuned" flavor with noticeably better prompt following and composition.
• Alpha-VLLM/Lumina-Next-T2I - the base text-to-image checkpoint. You'd pick it for experimentation or if you specifically want the raw behavior; the README lists both, and the default is default for a reason.

precision is bf16 (default, ~4GB, needs an Ampere-or-newer GPU) or fp32 (double the memory, for older cards). Same story as the Gemma loader: bf16 unless your GPU can't do it.

Output: lumina_model, which goes straight into LuminaT2ISampler. You never touch the weights after this - the sampler does all the real work.

How it loads

The interesting mechanism detail: the node builds the model with init_empty_weights (that's the accelerate dependency in the requirements) and then streams the safetensors weights onto it device-by-device. That's why the first load can look like it's hanging - it's reading a multi-gigabyte file. It also means the node can load onto an offload device and only move to the GPU when the sampler actually runs.

Installing the pack

ComfyUI Manager (search "ComfyUI-LuminaWrapper") or the old-fashioned way:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-LuminaWrapper
pip install -r requirements.txt

Restart, then let the node autodownload the model on first run (it's a few gigabytes, so go make tea). Optional but recommended: flash_attn. Without it, attention falls back to torch SDP, which the README warns is at least twice as slow and more VRAM-hungry. Windows users should grab a prebuilt wheel from the bdashore3 flash-attention releases instead of compiling.

What to expect, honestly

Lumina-Next is a niche, interesting model, not a daily driver. When it dropped in mid-2024 people praised its "very artistic" bent and decent prompt adherence, but it was trained on a heavy dose of synthetic data, which shows up as scraggly background details and unreliable humans. It's a model you reach for when you want that painterly, anything-goes look - not when you need a clean photoreal render. The one thing nobody complains about is the VRAM: two 2B models in bf16 is around 8GB total, so a mid-range card handles it.

And if you're wondering who to blame for this wrapper: kijai, the same person who ships KJNodes and the Wan/LTX wrappers, and who had the ComfyUI nodes up almost before the model's own authors finished the announcement thread.