ReSD35PatcherAdvanced

The block-selectable sibling of ReSD35Patcher. Same job - patching an SD3.5 model so RES4LYF's style-transfer and dual-conditioning system can operate on it - but with control over which of the model's transformer blocks the style guide is allowed to touch, instead of an all-or-nothing switch.

Why this one only has a single block field

SD3.5's MMDiT architecture doesn't split into separate "double-stream" and "single-stream" stages the way Flux, Chroma, and LTXV do - its transformer runs as one stack of joint blocks throughout, processing image and text tokens together at every layer. That's why this node exposes one joint_blocks field instead of the pair you'd see on the Flux or Chroma advanced patchers. It's the same idea, just matched to a different architecture: narrow the range to control whether the style signal leans compositional (earlier blocks) or textural (later blocks).

The inputs and outputs that matter

• model (MODEL, required) - your loaded SD3.5 model.
• joint_blocks (STRING, default "all") - comma-separated indices or ranges of SD3.5's joint transformer blocks to patch. Leave at "all" for the equivalent of the plain ReSD35Patcher.
• style_dtype (default float64) - precision for the style computation, independent of your model's own loaded dtype.
• enable (BOOLEAN, default true) - bypass switch.

Output: a patched model.

How to install it

Ships with RES4LYF, no separate step. 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 weights - Large, Large Turbo, or Medium, all share the joint-block architecture this patcher targets. RES4LYF's SD35Loader is a convenient way to get model, CLIP, and VAE loaded together if you don't already have that part of the graph set up.

Common issues & troubleshooting

Unsure where to start with joint_blocks. Leave it at "all", confirm the style effect matches what the plain patcher gives you, then narrow the range one step at a time on a fixed seed. There's no published guide to which specific block ranges do what on SD3.5 - this is genuinely trial and error, same as every advanced patcher in this pack.

A narrowed range appears to do nothing. This is a free-text field with no validation against the model's actual block count - a typo or out-of-range index fails silently. Rule that out before concluding the block narrowing itself isn't working.

VRAM pressure. SD3.5 Large in particular wants around 24GB at full precision; that's the base model's footprint, not this patcher's. If you're tight on memory, look at Medium instead, or quantization, before assuming style_dtype is your problem - it's a small addition relative to the full model pass.

You're reaching for this before confirming the basic patch works. Start with the plain ReSD35Patcher, confirm the style effect is present and roughly what you want, and only move to narrowing blocks here once you have a specific reason - too much style bleeding into detail, or not enough composition influence - to fix.