Skip to main content

​
Local models

Those are training peaks at 512px. Training uses less VRAM than generating, and a LoRA trained at 512 applies at any generation resolution. The full benchmark matrix is in TRAINING.md.
  • Z-Image Turbo is the low-VRAM place to start. It’s distilled to run without CFG, so 1024×1024 fits in about 11.5GB.
  • Krea 2 is a 12.9B model with an undistilled RAW base and an 8-step Turbo build. Generating at 1024 peaks near 36GB, so plan on a 40GB card.
  • FLUX.2 takes several reference images, and the prompt addresses them by position. One node covers klein 4B and 9B, their Base builds, and dev.
  • MiniMax H3 generates video and its soundtrack in one pass: 24fps, 5 to 15 seconds. Plan on 64GB of system RAM.
  • LTX-2.5 is a 22B video-with-sound model: 24fps, 1 to 20 seconds. Each node has a Fast mode and a Quality mode. It’s gated: accept the LTX-2 Community License on Hugging Face first.

​
Get the model files

Each model node has a popup that downloads the diffusion model, VAE and text encoder, with visible progress. Nothing downloads in the background without you asking. To use files you already have, put them under core/models/:
ComfyUI int8_convrot and int8_tensorwise files load unchanged and stay int8 in VRAM. A file the node can’t read is listed with the reason.

​
Hosted models

Add a Generate node and pick a model. The first provider is fal: FLUX.2, GPT Image 2, Nano Banana, Seedance, MiniMax H3, LTX and more. Add your fal key in Settings. It stays on your machine, and you pay fal per render.

​
Licences

Model weights carry their own licences. FLUX.2 klein 4B is Apache 2.0, while FLUX.2 dev, the 9B builds and FLUX.1 dev are non-commercial, and a LoRA trained on FLUX.1 dev inherits that. Krea 2, MiniMax H3 and LTX-2.5 have their own community licences. Check the licence on each model page before commercial use.