pipeLoaderSDXL

SDXL isn't one model with one text encoder - it's a base plus a refiner, and its CLIP is really two encoders: CLIP-L for "local" prompt detail and the OpenCLIP bigG for "global" style and composition. Most loaders hide that. tinyterra's pipeLoaderSDXL v2 (ttN pipeLoaderSDXL_v2) is the version that doesn't: it gives you separate positive_g / positive_l and negative_g / negative_l boxes, loads the base and refiner, applies a stack of LoRAs, and produces an sdxl_pipe ready for pipeKSamplerSDXL.

If you're doing SDXL in tinyterra, this is the loader you want. It's the current-generation version of the legacy pipeLoaderSDXL, and the big upgrade is exactly that split: in v1 you got one prompt string; here you can put the subject in _l and the style in _g the way the model was designed to be fed.

The inputs that matter

• ckpt_name (+ config_name if you need a non-default config), vae_name ("Baked VAE"), clip_skip (default -2).
• positive_g / positive_l and negative_g / negative_l - the four prompt boxes. A rough rule: global holds style, mood and medium; local holds the concrete subjects and details. The optional prepend_positive_g/l and prepend_negative_g/l inputs let you bolt fixed tags on without editing the boxes.
• loras - a multiline LoRA stack string (the pack's autocomplete helps here). Multiple LoRAs with model_strength and clip_strength each, in one text field - cleaner than ten separate slots, once you get used to it. You can also feed a LORA_STACK in via optional_lora_stack instead, plus optional_controlnet_stack.
• refiner_ckpt_name (+ refiner_config_name) - the refiner. Set to "None" if you're skipping it.
• The aspect/size block, which is where SDXL gets fiddly:
  • empty_latent_aspect - presets like "1024 x 1024 [S] 1:1", "640 x 1536 [P] 9:21" etc., which set empty_latent_width/height for you.
  • conditioning_aspect, conditioning_width/height, crop_width/height, target_aspect, target_width/height - this is SDXL's crop/target conditioning. SDXL conditions on the crop position and target size, and getting these wrong is why SDXL images come out squished or off-center. These fields let you control exactly how the conditioning is scaled to the latent.
• positive_ascore / negative_ascore - aesthetic-score conditioning for the refiner (defaults 6 and 2).
• batch_size, seed.

Outputs

sdxl_pipe plus the full breakdown: model, positive, negative, vae, clip, refiner_model, refiner_positive, refiner_negative, refiner_clip, latent, seed, and the handy width, height, pos_string, neg_string outputs for driving other nodes or filename prefixes.

Install and gotchas

One-time pack install - ComfyUI Manager → search ComfyUI_tinyterraNodes, or clone into ComfyUI/custom_nodes, restart. No extra dependencies; no bundled models.

The usual traps: leave the refiner "None" unless you've actually loaded one (a refiner with nothing behind it errors); and if outputs look slightly wrong, check the conditioning-aspect block before blaming the sampler - mis-set conditioning size is the classic SDXL complaint. And the honest caveat for the whole pack: this loader is SDXL-shaped. GGUF Flux and the newest architectures aren't what tinyterra's pipe nodes were built for, so for those you'll be using native nodes.