Given reference images, the i2L v2 hypernetwork predicts a LoRA in a single forward pass (no per-style training), then applies it to the Z-Image generator. This package targets the v2 model on ModelScope: DiffSynth-Studio/ZImage-i2L-v2.

v2 uses the new Diffusion Templates API (TemplatePipeline / call_single_side), which is different from v1's ZImageUnit_Image2LoRAEncode/Decode. Existing v1 community nodes do not run v2 by swapping a model id.

Nodes

| Node | In | Out | What it does | |---|---|---|---| | Loader | device, dtype, (cache dir) | pipe, template | Loads base Z-Image pipeline + i2L v2 template once. Models auto-download from ModelScope. | | Extract LoRA | template, IMAGE | lora | template.call_single_side(...) → predicted LoRA (deterministic; no seed). | | Save LoRA | lora, filename | filename | Writes .safetensors into models/loras. | | Generate | pipe, template, IMAGE, prompt, seed, cfg, steps, sigma_shift | IMAGE | Full styled image with asymmetric CFG (auto gray-image negative branch). |

Two usage patterns:

Loader → Extract → Save — get a reusable LoRA file.
Loader → Generate — best-quality styled image in one shot (keeps asymmetric CFG).

Requirements / setup (on the CUDA box, e.g. RunPod RTX 4090)

A 4090 (Ada, sm_89, 24 GB) works with mainstream stable PyTorch (cu121/cu124) — any current ComfyUI pod image already ships a compatible build. 24 GB is the known-good baseline; if you hit OOM, enable CPU offload. (On newer Blackwell cards like the 5090, you'd instead need a cu128+ PyTorch build.) Verify your setup with the env check below.

# 1. Install DiffSynth-Studio from git (NOT PyPI — v2 needs diffsynth.diffusion.template)
git clone https://github.com/modelscope/DiffSynth-Studio.git
cd DiffSynth-Studio && pip install -e . && cd ..

# 2. This node's light deps
pip install -r ComfyUI/custom_nodes/ComfyUI_ZImageI2L_v2/requirements.txt

# 3. (optional) point the model cache somewhere with space
export MODELSCOPE_CACHE=/workspace/modelscope_cache

# 4. Env check — must print True and import cleanly
python -c "import torch; print('cuda', torch.cuda.is_available(), torch.version.cuda); \
from diffsynth.diffusion.template import TemplatePipeline; print('v2 API OK')"

Then restart ComfyUI. First Loader run downloads the models (Z-Image + Z-Image-Turbo encoders/VAE + ZImage-i2L-v2 ≈ tens of GB, one-time) into the ModelScope cache.

Notes / known gaps

device defaults to cuda; mps/cpu are exposed but the upstream pipeline is CUDA-oriented.
sigma_shift defaults to 8 (matches the v1 i2L example); set 0 to omit the kwarg.
Out of scope (for now): multi-style fusion UI, ControlNet/inpaint composition, CPU-offload tuning.

Credits

Built on DiffSynth-Studio and Tongyi-MAI's Z-Image. Paper: Compressing Image Style Training into a Single Model Forward (arXiv 2606.13809). Apache-2.0.