SAX Loader

SAX Loader is where every SAX_Bridge workflow starts, and it's the node that makes the whole pack's promise real. Instead of three separate loaders for your checkpoint, your VAE, and your LoRA, plus a separate node for seed, steps, CFG, sampler, scheduler, and resolution, this one node does all of it and hands you a single PIPE that carries everything downstream. The pack calls it "CSB Loader (V3)" internally - it's the third generation of this design - and it's the most comfortable seat in the house: load once, wire one thing, done.

What it loads and sets

All in one node, in one pass:

• ckpt_name - your checkpoint. vae_name - its VAE, or baked_vae to use the one inside the checkpoint.
• lora_name + lora_model_strength - an optional LoRA applied right at load time. The strength slider drives both the model and the CLIP at once (one value, two targets - see the pack docs), which is a deliberate simplification.
• clip_skip - number of CLIP layers to skip, on the model. Only meaningful for CLIP-based SD/SDXL checkpoints; on newer LLM-encoded models there's no CLIP to skip.
• v_pred - a boolean that flips the pipeline into V-prediction mode (the pack applies V_PREDICTION + ZSNR sampling when you turn it on). For the V-pred models that want it.
• The sampler set that becomes your defaults: seed, steps, cfg, sampler_name, scheduler_name, denoise.
• width, height (multiples of 8), batch_size.

What comes out

Two outputs: PIPE and SEED (the seed as a plain integer, handy for driving filename templates or a second sampler). The pipe contains model, clip, vae, the empty latent (samples), the seed, and a loader_settings dict holding steps/CFG/sampler/scheduler/denoise. Positive and negative conditioning start as None - a SAX Prompt node fills those in - and images starts empty until a sampler decodes. The point of that structure: every downstream SAX node reads its settings from the pipe, so your sampler, upscaler, and detailer all inherit the same steps and CFG without you re-entering them. That's the "fewer nodes, fewer mistakes" argument in action.

Where it fits

The canonical flow the README pushes: SAX Loader → SAX Prompt → SAX KSampler → SAX Upscaler → SAX Detailer → output. Every one of those downstream nodes pulls what it needs off the pipe, and steps_override/cfg_override on the detailers let you punch a different value into a single stage without touching the global defaults. If you're moving from a stock ComfyUI graph, the mental model swap is: the loader now owns the sampler settings too, and the KSampler is nearly a pass-through that reads them.

Installing

It ships with the pack, no extra dependencies:

cd ComfyUI/custom_nodes
git clone https://github.com/so16tm/SAX_Bridge

Restart and you'll find it under SAX/Bridge/Loader. Two things worth knowing. First, if your checkpoint is a split model (like Anima) with no baked CLIP/VAE, reach for the SAX Diffusion Loader instead - that one loads UNET/CLIP/VAE from separate folders. Second, the clip_skip default of -1 and CFG default of 8 are SDXL-era habits; if your checkpoint is a 2026 flow-matching model that wants CFG 1, set cfg accordingly and ignore clip_skip entirely. The loader won't fight you, but it won't know your model's generation either.