ReHiDreamPatcher

HiDream (specifically HiDream-I1, the 17B MIT-licensed model from April 2025 - not to be confused with the later, unrelated HiDream-O1) had a short but real moment: best-in-class prompt adherence for an open model at the time, driven by a dual text encoder stack including Llama-3.1-8B. It also faded fast, mostly because 17B parameters is a lot to carry on consumer hardware and the community spent its first month just trying to quantize it down to something an RTX card could run. RES4LYF's README lists HiDream among its supported style-transfer architectures regardless - the pack's whole design goal is broad model coverage, not chasing whatever's trending this month - and ReHiDreamPatcher is the node that wires HiDream into that system.

What it does

Same job as every Re*Patcher node in this pack: it wraps your loaded model so RES4LYF's style-guide and dual-conditioning nodes can inject their extra guidance signal correctly, which requires knowing HiDream's specific transformer internals rather than assuming a generic transformer layout. This is the plain, on/off version - flip it on, it patches the whole model; flip it off, it's a no-op passthrough. If you need finer control over which transformer blocks the style signal reaches, that's ReHiDreamPatcherAdvanced.

The inputs and outputs that matter

• model (MODEL, required) - your loaded HiDream model.
• style_dtype (default float64) - precision for the internal style computation specifically, independent of whatever dtype your (likely already VRAM-hungry) HiDream weights are loaded at.
• enable (BOOLEAN, default true) - bypass switch to A/B the patch's effect without rewiring the graph.

Output: a patched model, wired onward to your sampler like normal.

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

(swap in your portable install's embedded pip.exe on Windows). You'll need HiDream's weights yourself - this node patches an already-loaded model, it doesn't download anything. Be realistic about the hardware ask: 17B parameters plus a Llama-8B text encoder is a genuinely heavy stack, and most of the community's practical experience with it in 2025 was about getting it to fit at all, via aggressive quantization, rather than about fine-tuning style transfer on it.

Common issues & troubleshooting

Out of memory before you even get to the style patch. This is almost certainly HiDream itself, not this node - it was, and remains, one of the heaviest open models people ran locally. Quantize the diffusion weights and the text encoder before assuming the patcher's float64 setting is the problem; the style computation touches a small slice of the network compared to the full model pass.

No visible effect. Double check enable is actually on, and that your workflow is genuinely feeding a style guide or dual-conditioning setup downstream - this patch only matters to nodes that use it.

You're not sure HiDream is worth the trouble for this. That's a fair question to ask honestly. If you already have a HiDream workflow running for its prompt-adherence strength and want style transfer on top of it, this is the right tool. If you're starting from scratch, most people doing style-transfer work in 2026 are reaching for Flux, Chroma, or newer models first and HiDream is the heavier, more niche option among RES4LYF's supported architectures.