ReHiDreamPatcherAdvanced

The advanced sibling of ReHiDreamPatcher: same job - hooking RES4LYF's style-transfer and dual-conditioning system into a HiDream model - but instead of a single on/off toggle, you get to choose exactly which of HiDream's transformer blocks the style guide is allowed to touch. Reach for this once the default "patch everything" behavior is close but not quite giving you the balance of composition vs. detail you want.

Why block selection matters here specifically

HiDream-I1 is a big, dense 17B transformer, and like other diffusion transformers, its early blocks tend to shape broad composition while later blocks refine texture and fine detail. Restricting RES4LYF's style-guide injection to a subset of blocks lets you push the style effect toward one end or the other - early double-stream blocks for a compositional style influence, later single-stream blocks for something closer to a texture/finish pass. There's no shortcut map for exactly which block ranges do what on HiDream specifically; this is trial-and-error territory, same as the advanced patchers for every other architecture in the pack.

The inputs and outputs that matter

• model (MODEL, required) - your loaded HiDream model.
• double_stream_blocks (STRING, default "all") - comma-separated indices or ranges of HiDream's double-stream blocks to patch.
• single_stream_blocks (STRING, default "all") - same, for the single-stream blocks.
• 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 HiDream's weights yourself. Go in expecting a heavy setup - HiDream-I1 at 17B parameters plus its Llama-8B text encoder was, and still is, one of the more VRAM-demanding options in RES4LYF's supported model list, and that's true whether or not you're using this patcher.

Common issues & troubleshooting

You don't know where to start. Leave both block fields at "all" first, confirm the effect matches what the plain ReHiDreamPatcher would give you, then narrow one field at a time on a fixed seed and converging sampler so you're actually comparing the block change and nothing else.

A narrowed range does nothing. These are free-text fields with no index validation - an out-of-range or malformed block reference fails silently rather than erroring. If narrowing seems to have no effect, double-check the string first before concluding the technique doesn't apply.

Memory pressure before you even reach the patching step. That's HiDream itself, not this node. Sort out quantization and encoder precision for the base model before spending time tuning block ranges - a model that won't load at all makes the block selection moot.

Reaching for this reflexively. Most people should start with the plain ReHiDreamPatcher and only move to the advanced version once they've confirmed the default style patch is doing something worth refining. Narrowing blocks on a model you haven't validated the basic patch on yet is optimizing a step you haven't confirmed works.