Apply Super Adapter (Flux/SDXL)

It's a LoRA loader with the "low-rank" part taken out

ApplySuperAdapter is the entire ComfyUI-SuperAdapter pack, and it does one thing: load a ~1.3 GB safetensors file of weight deltas, add them straight onto your diffusion model, and hand the patched model back. No API, no key, no trigger word, no CLIP at all. You drop it between your model loader and the KSampler, set one slider, and sample.

The name oversells the mechanism - the clever part is in how the weights were made, not how the node applies them. The file, "Super Adapt" (SAF_ Flux-krea.safetensors), is a real fine-tune of FLUX.1-Krea-dev. The author says they tested layers one by one until double-blocks 17 and 18 - the last pair of Flux's dbl_blocks - gave the look they wanted, and shipped only that delta. That's why it's a gigabyte instead of a 12 GB checkpoint: it's the model's change, not the model.

Their own pitch, from the model card: it doesn't touch skin tone, it "amplifies the signature plastic-like tone present in FLUX," and it pushes realism. If your prompt is weak, beware - it can veer into oil-painting territory. That last warning is not boilerplate; the author put it there for a reason.

How it actually works

The whole node is about 40 lines. It loads the safetensors, walks every key (prepending diffusion_model. where missing, which is how ComfyUI names UNet weights), and calls model.add_patches(...) - the exact code path ComfyUI's built-in LoRA loader uses. The difference is what's inside the file: a LoRA stores small low-rank factors, this stores real per-layer deltas, so the injection is full-rank. Think of it as a hidden fine-tune you can dial in at runtime instead of committing to at download time.

Two behaviors worth knowing:

• Strength is baked into the delta. The node multiplies every weight by your strength before patching, so 0.5 genuinely halves the effect. And strength == 0.0 short-circuits - it returns the model untouched, no patching at all.
• The slider goes to −10…10, not the usual 0…2. Negative inverts the delta. The wide range is the author telling you to experiment.

The inputs that matter

Three inputs, one output - and only two of the inputs are things you'll actually touch:

• adapter_name - a dropdown of everything in ComfyUI/models/super_adapter/. The pack creates that folder automatically the first time it loads. One real quirk: the example workflow on Hugging Face expects a file named super_adap.safetensors, but the repo actually ships SAF_ Flux-krea.safetensors (space included). Rename it to whatever you like, or just pick it from the dropdown and move on.
• strength - start at 0.5, not 1.0. The author's own oil-painting warning is your cue to ease in.

The model input takes any MODEL and the single MODEL output feeds the KSampler. The bundled Basic_workflow_SAF.json wires it exactly that way: UNETLoader → ApplySuperAdapter → KSampler.

Installing it

Either search "SuperAdapter" in ComfyUI Manager, or clone it:

cd ComfyUI/custom_nodes
git clone https://github.com/kpsss34/ComfyUI-SuperAdapter

Restart ComfyUI. Then download the model from kpsss34/SAF_Super-Adapt.1 and drop it in ComfyUI/models/super_adapter/ - nothing downloads itself. No Python dependencies beyond what ComfyUI already ships (torch + safetensors). The optional HDR VAE linked in the README is a separate, gated download, so you'll need a logged-in Hugging Face account for that one; it's not required for the adapter.

Gotchas worth knowing

• Silent mismatch is the big trap. The node never checks whether the adapter's keys exist in the model you feed it. Wrong architecture and the patches are silently ignored - an SDXL adapter on Flux, or vice versa, is a no-op that looks like it ran. This file is tuned for FLUX.1-Krea-dev specifically.
• Empty dropdown = you skipped the download. The install step alone gives you nothing to select.
• Oil-painting artifacts are the documented failure mode at low prompt quality. Lower strength or rewrite the prompt before blaming the node.

One honest note: this pack has essentially zero community footprint - no reddit chatter, a couple of likes on the model, a single author. It works fine, but it's a personal project, so read the source (it's tiny) before building a production workflow around it. That's good practice for any custom node this obscure.