<h1 id="rebels-nl-diffusion-image-ggufs">Rebels NL-Diffusion-Image GGUFs</h1><h1 id="comfyui_rebels_nld-custom-nodeswork-in-progress.-expect-bad-outputs-until-i-explicitly-change-this-description.">ComfyUI_Rebels_NLD Custom Nodes WORK IN PROGRESS. EXPECT BAD OUTPUTS UNTIL I EXPLICITLY CHANGE THIS DESCRIPTION.</h1>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):<a target="_blank" rel="ugc" href="https://huggingface.co/realrebelai/NL-Diffusion-Image_GGUF/tree/main">https://huggingface.co/realrebelai/NL-Diffusion-Image_GGUF/tree/main</a> CUSTOM NODES (REQUIRED):<a target="_blank" rel="ugc" href="https://github.com/RealRebelAI/ComfyUI_Rebels_NLD">https://github.com/RealRebelAI/ComfyUI_Rebels_NLD</a>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.<h2 id="install">Install</h2><ol><li>Clone into <code>ComfyUI/custom_nodes/</code>.</li><li>Requires the city96 ComfyUI-GGUF fork in the same <code>custom_nodes/</code> folder (used for dequant).</li><li>Put the model files (dropdown-selected, no paths):<ul><li>dLM GGUF → <code>ComfyUI/models/unet/</code></li><li>vqvae (bf16 <code>.safetensors</code>) → <code>ComfyUI/models/vae/</code></li></ul></li><li>The config/tokenizer/modeling code ships in <code>model_assets/</code></li></ol><h2 id="important!">IMPORTANT!</h2><ol><li>model.safetensors file MUST go in "custom_nodes\ComfyUI_Rebels_NLD\model_assets\emu3_vqvae"</li></ol><a target="_blank" rel="ugc" href="https://huggingface.co/nvidia/NL-Diffusion-Image/blob/main/emu3_vqvae/model.safetensors">https://huggingface.co/nvidia/NL-Diffusion-Image/blob/main/emu3_vqvae/model.safetensors</a><h2 id="nodes">Nodes</h2><ul><li>NL-Diffusion dLM Loader (GGUF) — pick <code>gguf_name</code> and <code>vqvae_name</code> from dropdowns, choose device.</li><li>NL-Diffusion Text to Image — prompt, size, steps, guidance, temperature, seed → IMAGE.</li></ul><h2 id="notes">Notes</h2><ul><li>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.</li><li>Vocab embeddings use row-gather dequant, so the 131k-row tensors never fully materialize.</li><li>Vision-tower (image-understanding / edit) weights are left on meta and not needed for t2i.</li></ul><h2 id="license">License</h2>Model is under the NVIDIA One-Way Noncommercial License (research/development only). Quants inherit those terms — publish as <code>license: other</code> with the upstream terms linked.

Rebels NL-Diffusion-Image (GGUF)

nvidia sucks but their model isnt that bad.

Files with SHA256: 7166f820e545b8aaf8a5de7fdec2eee66256956e802ec5a9bf4c57c30a18fa34