SD35Loader

SD3.5 splits into more moving pieces than an SDXL checkpoint does - the diffusion model, up to three CLIP text encoders (SD3.5 supports a triple encoder stack), and a VAE, any of which can come from separate files depending on how you sourced your setup. SD35Loader collapses that into one node: pick your diffusion model, your weight precision, up to three CLIP files (or let it use whatever's bundled in a checkpoint), and a VAE, and get back a ready-to-use MODEL, CLIP, and VAE in one place instead of wiring three or four separate loader nodes together.

Why this exists as its own node

Stock ComfyUI's loaders assume a fairly standard shape - one CheckpointLoaderSimple, or a UNETLoader plus a DualCLIPLoader. SD3.5's actual release structure doesn't map cleanly onto either: Stability shipped the model in multiple weight-precision variants (including fp8 quantized options for lower-VRAM cards), and depending on which files you downloaded, your CLIP and VAE might be baked into a checkpoint or might be separate .safetensors files you need to point at explicitly. This loader is built for that reality - it exposes a weight_dtype choice for the diffusion model directly, and gives every CLIP slot and the VAE slot an option to fall back to whatever the checkpoint ships (.use_ckpt_clip, .use_ckpt_vae) instead of forcing a separate file.

The inputs and outputs that matter

• model_name - your SD3.5 diffusion model file, picked from whatever's in your models folder.
• weight_dtype - precision for the loaded model: default, or one of the fp8 variants (fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2) if you're on a lower-VRAM card and have an fp8 build of the model available.
• clip_name1 - your primary CLIP encoder, or .use_ckpt_clip to pull it from the checkpoint instead of a separate file.
• clip_name2_opt / clip_name3_opt - SD3.5's additional encoder slots (it supports up to three), each defaulting to .none if you're not using them. Leave these off unless you specifically have the extra encoder files and want the full triple-encoder setup.
• vae_name - your VAE, or .use_ckpt_vae to use the checkpoint's own. The dropdown also lists the small taesd-family approximate VAEs (taesd, taesdxl, taesd3, taef1) for fast previews - these are lightweight preview decoders, not full-quality output VAEs, so don't leave one selected for your final render.

Outputs: model, clip, and vae - wire each into the corresponding input on your sampler, conditioning, and VAE decode nodes exactly as you would from any other loader.

How to install it

Comes with RES4LYF - no separate install. Via ComfyUI Manager: search "RES4LYF", install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ClownsharkBatwing/RES4LYF/
cd RES4LYF
pip install -r requirements.txt

You still need to download SD3.5's actual weights yourself - RES4LYF doesn't bundle any model files. Stability's official Hugging Face releases (Large, Large Turbo, Medium) are the source; place the diffusion model, CLIP, and VAE files in their usual ComfyUI model folders so this loader's dropdowns can see them.

Common issues & troubleshooting

A dropdown is empty or missing your file. These lists are populated from your actual model folders at ComfyUI startup - if you just downloaded a file, restart ComfyUI (or refresh the node) before assuming the loader is broken.

You left a taesd-family option selected in vae_name and your final image looks soft or wrong. Those are the fast-preview approximate VAEs, meant for live preview during generation, not for final decode quality. Switch to the real SD3.5 VAE (or .use_ckpt_vae) for anything you're keeping.

Using fp8 weight_dtype and seeing quality loss you didn't expect. SD3.5 Large in particular is a large model (~24GB at full precision); fp8 is a reasonable trade for fitting it on a smaller card, but if quality matters more than VRAM headroom, try default first and only step down if you actually need to.

CLIP or VAE mismatch errors. Mixing an SD3.5 CLIP/VAE with a non-SD3.5 diffusion model (or vice versa) will produce shape errors or garbage output - SD3.5's components aren't drop-in compatible with SDXL, Flux, or other architectures despite superficially similar file formats.