JoyLTX Load Model

The Joy-LTX 2.5 model ships in two families: GGUF quants for RTX 30/40 cards and comfy-native .safetensors files for RTX 50. ComfyUI's stock loader won't touch a .gguf, and the GGUF loader won't look at a native file, which means most workflows need to know which card you own. JoyLTX Load Model is the pack's way of not caring: one node that routes .gguf files to ComfyUI-GGUF's loader and everything else to the stock Load Diffusion Model. Pick your file, done.

The model_name dropdown lists every file in models/diffusion_models plus models/unet, including .gguf files buried in subfolders that the stock loader never enumerates. If you're on a 24 GB card, that's the ...070T30-DiT-Q5_K_M.gguf (or the ...mix4x8-17.0GB.safetensors on a 50-series); 16 GB cards take the Q4_K_S or w4a8 builds, 32 GB the Q8_0 or int8. The README's pick-by-card table is the ground truth here.

Two behaviors make it friendlier than a plain loader. First, it resolves by file name: save a workflow with a bare filename, file the model into a subfolder later, and the loader finds it by unique basename instead of erroring on the old path. Second, when you pick a .gguf and ComfyUI-GGUF isn't installed, you get a clear "install ComfyUI-GGUF" error rather than a cryptic import failure. The optional weight_dtype input only applies to .safetensors files - it's the usual fp8 menu, and you can ignore it if the dropdown defaults are fine.

One thing worth knowing before you queue: the loader watches which backend comfy-kitchen has live, and if only its "eager" backend is running it prints a warning that comfy-native quantized files (int8/w4a8) dequantize every step and run about 2x slower. GGUF is unaffected. It's a hint, not a blocker - if you're on comfy-native weights and the renders feel slow, that's the message to search for. Add --enable-triton-backend to your launch line or switch to the GGUF build.

Output is a plain MODEL, wired into the sampler or the LoRA stack like any other loader. There's nothing clever in the render itself - this node is all about not making you manage two loader types and two file paths by hand. For a pack whose whole pitch is "install in five minutes and run," that's the right kind of boring.