pipeLoaderSDXL v1 (Legacy)

The SDXL-era tinyterra loader was a big node and it earned every input on it. pipeLoaderSDXL (ttN pipeLoaderSDXL, v1, Legacy) loaded your base checkpoint and your refiner, applied up to two LoRAs to each, took positive and negative prompts with A1111-style weight interpretation, set the latent size, and packed it all into an sdxl_pipe - one wire, one node, everything an SDXL sampler needed.

It's legacy now, but it's the ancestor of the v2 SDXL loader in the pack, and it still runs in any old workflow that references it. If you're starting fresh you want the v2, for one concrete reason: SDXL is a two-text-encoder model, and the v1 treated the prompt as a single string, while v2 gives you separate global (_g) and local (_l) prompts the way SDXL actually wants them. The v1 uses the vendored ADV_CLIP_emb encoding to squeeze a single prompt through both encoders, which works fine - it's just less precise about which half of the prompt hits which encoder.

What it takes in

• ckpt_name, vae_name ("Baked VAE" by default), clip_skip (default -2, the SDXL sweet spot).
• lora1_name/lora1_model_strength/lora1_clip_strength and the same trio for lora2 - base model LoRAs.
• refiner_ckpt_name, refiner_vae_name, plus refiner_lora1_* and refiner_lora2_* - the refiner gets its own LoRA slots. "None" means skipped.
• positive / negative with token_normalization and weight_interpretation per prompt.
• empty_latent_width, empty_latent_height, batch_size, seed.

What comes out

sdxl_pipe plus the full breakdown: model, positive, negative, vae, clip, refiner_model, refiner_positive, refiner_negative, refiner_vae, refiner_clip, latent, and seed. Note the double set - base and refiner each carry their own model, VAE, conditioning and CLIP, which is exactly what the paired pipeKSamplerSDXL expects to run its base+refiner pass.

Install and gotchas

One-time pack install - ComfyUI Manager → search ComfyUI_tinyterraNodes, or git clone https://github.com/TinyTerra/ComfyUI_tinyterraNodes.git into ComfyUI/custom_nodes, restart. No extra dependencies; no bundled models (the refiner is a checkpoint you download separately).

The gotchas are the legacy-family ones: it's SDXL-only by construction, and the single-string prompt is the biggest reason to migrate to v2 if you care about fine prompt control (or to modern native SDXL encoding if you've outgrown the pipe system entirely). And like the rest of the pack's v1 nodes, don't expect newer architectures here - this is the era where the pack's community-reported weak spot (no GGUF/Flux support) starts showing.