Rebels NL-Diffusion-Image GGUFs

ComfyUI_Rebels_NLD Custom Nodes


WORK IN PROGRESS. EXPECT BAD OUTPUTS UNTIL I EXPLICITLY CHANGE THIS DESCRIPTION.

GGUF loader + text-to-image nodes for NVIDIA NL-Diffusion-Image (masked discrete diffusion LM + IBQ VQ decoder) on consumer hardware. By RealRebelAI.

MODELS (Transformer, Encoder, VAE):

https://huggingface.co/realrebelai/NL-Diffusion-Image_GGUF/tree/main

CUSTOM NODES (REQUIRED):

https://github.com/RealRebelAI/ComfyUI_Rebels_NLD

NODES ARE OPERATIONAL BUT SLOW. Flash ATtention helps speed up significantly. Currently working on patches for speed ups. they will run in their current state but i recommend git pulling frequently.

Install

1. Clone into ComfyUI/custom_nodes/.

2. Requires the city96 ComfyUI-GGUF fork in the same custom_nodes/ folder (used for dequant).

3. Put the model files (dropdown-selected, no paths):

   • dLM GGUF → ComfyUI/models/unet/

   • vqvae (bf16 .safetensors) → ComfyUI/models/vae/

4. The config/tokenizer/modeling code ships in model_assets/

IMPORTANT!

1. model.safetensors file MUST go in "custom_nodes\ComfyUI_Rebels_NLD\model_assets\emu3_vqvae"

https://huggingface.co/nvidia/NL-Diffusion-Image/blob/main/emu3_vqvae/model.safetensors

Nodes

• NL-Diffusion dLM Loader (GGUF) — pick gguf_name and vqvae_name from dropdowns, choose device.

• NL-Diffusion Text to Image — prompt, size, steps, guidance, temperature, seed → IMAGE.

Notes

• The dLM generates discrete token indices; the vqvae decoder turns them into pixels. It is not a latent VAE — it loads through this pack, not ComfyUI's VAELoader.

• Vocab embeddings use row-gather dequant, so the 131k-row tensors never fully materialize.

• Vision-tower (image-understanding / edit) weights are left on meta and not needed for t2i.

License

Model is under the NVIDIA One-Way Noncommercial License (research/development only). Quants inherit those terms — publish as license: other with the upstream terms linked.