SP_SDLoader
SP_SDLoader — the classic checkpoint loader that needs zero extra packs
- sp_pipe
- model
- clip
- vae
- positive
- negative
- latent
Here's the thing that makes SP_SDLoader the odd one out in this pack: every other loader in SP-Nodes drags along dependencies - ComfyUI-GGUF here, a bitsandbytes pack there. This one uses only stock ComfyUI nodes. It loads an SD 1.5/SDXL checkpoint, optional LoRA, encodes your prompt, and hands you a ready-to-sample pipe, and it does it with nothing but what ships in ComfyUI itself. If you're on a plain SD checkpoint and you want the pack's "one node instead of six" convenience without installing anything else, this is the one.
It's a group node like its siblings - at execution time ComfyUI's GraphBuilder expands it into CheckpointLoaderSimple, an optional VAELoader and LoraLoader, a CLIPSetLastLayer, two CLIPTextEncode, and an EmptyLatentImage - so the single widget panel is really a stock SD rig with a bow on it. Outputs are sp_pipe, model, clip, vae, positive, negative, latent, feeding the pack's SP_KSampler through the SP_PIPE type or any ordinary sampler.
Inputs you'll actually set:
- ckpt_name - from
models/checkpoints, so SD 1.5 and SDXL merge files both work. This is the one input that's a checkpoint, not a separate unet. - vae_name -
Bakedor an external VAE frommodels/vae.Bakedis the pragmatic default: it uses whatever VAE is fused into the checkpoint, which is fine for the vast majority of SD 1.5/SDXL merges. - lora_name + lora_strength - optional, applied at strength to both model and clip.
- stop_at_clip_layer - the old clip-skip dial, default
-1(no skip). Some SD 1.5 checkpoints were trained expecting-2; SDXL should stay at-1because clip skip does nothing for it. If you don't know why you're changing it, leave it. - positive / negative, plus the resolution presets (31 of them;
width x height (custom)falls through toempty_latent_width/height).
A note for people coming from the newer models: this node is for the old family, where negative prompts and CFG 4–7 genuinely do what you expect - the guidance-distillation rules that make CFG 1 and no-negative the modern default don't apply here. That's exactly why this loader still makes sense: SD 1.5 and SDXL fine-tunes (Illustrious, NoobAI, Pony, WAI) are a huge chunk of what people actually run, and they behave like the docs tell you.
Install is the pack-wide recipe:
cd ComfyUI/custom_nodes
git clone https://github.com/bananasss00/ComfyUI-SP-Nodes
or ComfyUI Manager → SP-Nodes, then restart. On first launch the pack auto-installs nltk and pyyaml and downloads the punkt tokenizer - harmless background noise. Because SP_SDLoader stays on CheckpointLoaderSimple, it's also the most forgiving SP loader to start with if you're just trying the pack's group-node workflow: install the pack, drop in SP_SDLoader + SP_KSampler, and you've got a working SD rig with no missing-node surprises.
Inputs (10)
| Name | Type | Default | Description |
|---|---|---|---|
| ckpt_name | COMBO | 0 options: | |
| vae_name | COMBO | 2 options: Baked, pixel_space | |
| lora_name | COMBO | 1 options: None | |
| lora_strength | FLOAT | 1.00-100–100 | — |
| positive | STRING | — | |
| negative | STRING | — | |
| stop_at_clip_layer | INT | -1-24–-1 | — |
| resolution | COMBO | 1024 x 1024 | 31 options: width x height (custom), 512 x 512, 512 x 768, 576 x 1024, 768 x 512, 768 x 768, +25 |
| empty_latent_width | INT | 51264–16384 | — |
| empty_latent_height | INT | 51264–16384 | — |
Outputs (7)
| Name | Type | Description |
|---|---|---|
| sp_pipe | SP_PIPE | — |
| model | MODEL | — |
| clip | CLIP | — |
| vae | VAE | — |
| positive | CONDITIONING | — |
| negative | CONDITIONING | — |
| latent | LATENT | — |