Omini Kontext Split Pipeline Loader

The main Pipeline Loader downloads the whole diffusers model from the Hub in bf16, and on a 16GB card that's a wall. This node is the alternative: the Omini Kontext Split Pipeline Loader builds the same OMINI_KONTEXT_PIPELINE from local files - your own transformer, text encoders, and VAE - and it's the only node in the pack that talks GGUF. If you've been reading the pack's VRAM advice and thinking "great, but I don't have 24GB", this is your way in.

How it works

Instead of one from_pretrained call, it assembles the pipeline piece by piece. The transformer comes from ComfyUI's models/unet folder (safetensors or GGUF), the two text encoders - CLIP and T5 - come from models/clip, and the VAE from models/vae. It builds each component from bundled config files (shipped in the node's configs.zip, extracted on first run), loads the weights, then ties them together into the same FluxOminiKontextPipeline the regular loader produces. VAE slicing and tiling get enabled automatically.

Two details make this worth seeking out over the main loader:

• GGUF transformer support. If your transformer file ends in .gguf, it's loaded through diffusers' GGUFQuantizationConfig - a quantized Flux transformer, which is exactly the trick people use to get a 12B model running on 16GB.
• auto_t5_gguf (a boolean, default off). Flip it on and the T5 text encoder is pulled as a 4-bit GGUF (t5xxl_fp16-q4_0.gguf from the calcuis/kontext-gguf repo) instead of a full safetensors file. The T5 is a big chunk of the VRAM bill in these Flux-family pipelines, and q4 eats a fraction of it.

The fields it wants are the four file pickers - transformer_path, clip_path, t5_path, vae_path - which scan their folders at load time, plus optional lora_path and that auto_t5_gguf toggle. Output is the usual OMINI_KONTEXT_PIPELINE, so everything downstream (the Pipeline node, LoRA nodes, encoders) works exactly the same.

What you need to supply

This node assumes you already have the pieces: a Kontext-capable transformer, the CLIP and T5 encoders, and a Flux VAE, all in the right ComfyUI folders. That's more setup than the one-click main loader, and it's the honest trade - you're trading convenience for the ability to use files you already downloaded and quantized versions you wouldn't otherwise get. If you don't have the files, grab them from a FLUX.1-Kontext diffusers download and strip out the safetensors.

The catch

The pack is a wrapper, and the split loader is the wrapper's most ambitious node - it leans on bleeding-edge diffusers features like from_single_file with GGUF quantization, plus a slightly odd T5 fallback. It's genuinely useful for low-VRAM runs and it's the only path here that dodges the 23GB Hub download and the gated-model login. But it's also the most likely node in this pack to break when diffusers upstream moves, because it depends on those newest bits. If it misbehaves, the requirement is the same as everywhere in this pack: make sure you installed diffusers from git main, not a stable release.