ReSD35Patcher

SD3.5 was Stability AI's attempt to make good on SD3's disastrous launch, and it mostly worked as a technical fix - better anatomy, a genuinely more permissive license, three sizes to fit different hardware - even though it debuted below Flux on independent rankings and never fully won the community back. It's still a real, actively-used model, particularly SD3.5 Medium on modest hardware and SD3.5 Large for people who want a non-distilled, better-licensed alternative to Flux Dev. ReSD35Patcher is what lets RES4LYF's style-transfer and dual-guide system operate on it.

What it does

Same pattern as every Re*Patcher node in the pack: it wraps your loaded SD3.5 model so RES4LYF's style-guide machinery can inject its extra conditioning signal into the forward pass correctly, using SD3.5's own MMDiT transformer layout rather than assuming a generic one. This is the plain on/off version. SD3.5's transformer architecture doesn't split into the double-stream/single-stream layout you see in Flux-family models - its blocks are "joint" blocks throughout - which is why its advanced sibling exposes one block field instead of two; see ReSD35PatcherAdvanced for that control.

The inputs and outputs that matter

• model (MODEL, required) - your loaded SD3.5 model (Large, Large Turbo, or Medium - all share the same core architecture).
• style_dtype (default float64) - precision for the internal style computation, independent of whatever dtype your SD3.5 weights are running at.
• enable (BOOLEAN, default true) - bypass switch to compare patched vs. unpatched without rewiring.

Output: a patched model, wired onward to your sampler like any other model output.

How to install it

Comes bundled 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'll need your own SD3.5 checkpoint (this pack's SD35Loader node is one convenient way to load one alongside its CLIP and VAE), since this patcher operates on an already-loaded model rather than downloading anything itself.

Common issues & troubleshooting

No visible effect from toggling this. Confirm enable is on and that your workflow actually routes through a RES4LYF style-guide or dual-conditioning node downstream - the patch is inert if nothing consumes the signal it enables.

You're on SD3.5 Large and running low on VRAM. That's SD3.5 itself (roughly 24GB for Large at full precision), not this node. This patcher's float64 style computation is a small addition on top; sort out your base model quantization or offloading first if memory is tight.

You're unsure whether SD3.5 or Flux is the better base for style-transfer work right now. Both are reasonable choices and RES4LYF supports both - SD3.5's advantage is licensing and being an undistilled base model; Flux's is a much larger surrounding LoRA and tooling ecosystem. Neither is objectively wrong, and this patcher exists precisely so you're not forced to choose based on which one has style-transfer support.

Confusing this with a different Stability model. SD3.5 is not SD3 - SD3 was the version pulled from CivitAI entirely after its license fight and safety-tuning disaster. Make sure your checkpoint is actually the 3.5-family release before troubleshooting anything as an architecture mismatch.