Load InfuseNet

The heart of InfiniteYou is InfuseNet, and it is not a LoRA and not an IP-Adapter - it's a ControlNet rebuilt for FLUX's transformer architecture. Instead of a copy of a UNet's encoder blocks, it's a stack of small residual blocks that run alongside FLUX's double and single attention blocks, and it's what physically gets the identity into the generation: the "inject identity features via residual connections" line from the paper's abstract, made real. Load InfuseNet is the unglamorous node that puts that network on the table. It has exactly one input and one output, and it hides the pack's only real hardware decision.

The one input, and what those four names mean

controlnet_name, a dropdown with four choices:

• aes_stage2/infusenet_aes_fp8e4m3fn.safetensors - the default in every shipped workflow
• aes_stage2/infusenet_aes_bf16.safetensors
• sim_stage1/infusenet_sim_fp8e4m3fn.safetensors
• sim_stage1/infusenet_sim_bf16.safetensors

The stage names come straight from the paper's two-stage training recipe: sim_stage1 is the identity-similarity pretraining pass, aes_stage2 the final aesthetics-focused supervised fine-tune. The README is explicit that you should use the version corresponding to the image_proj_model.bin you loaded in ID Embedding Model Loader - mixing stages isn't the supported path, and aes_stage2 + aes_stage2 is what every example loads.

The real choice is bf16 versus fp8e4m3fn. Full-precision bf16 peaks around 43GB of VRAM; the fp8 quant runs in about 24GB. The example workflow ships on fp8, and on a 24GB card that's the only sane option. Got 48GB and want maximum fidelity? bf16 is there waiting.

How it works

On load it checks ComfyUI/models/infinite_you for the file, downloads it from Hugging Face if it's missing (it's a multi-GB pull the first time), then hands it to ComfyUI's own controlnet loader. InfuseNet subclasses ComfyUI's ControlNet class, which is the smart bit - everything downstream behaves like a normal controlnet. That's why strength, start/end percent, and the Apply node all feel familiar even though the identity math inside is new.

The single output, CONTROL_NET, feeds straight into Apply InfuseNet. There's nothing else to do with it; don't wire it into the sampler, that's not its job.

Installing

Standard pack install, and it's on the Comfy Registry so Manager works too:

cd ComfyUI/custom_nodes
git clone https://github.com/bytedance/ComfyUI_InfiniteYou.git
cd ComfyUI_InfiniteYou
pip install -r requirements.txt

Then restart, and be ready for the auto-download on first load.

Where people get burned

• It needs a recent ComfyUI. The pack reaches into comfy.ldm.flux and the newer controlnet internals. If the node errors out on import, update ComfyUI before you blame the pack.
• Don't assume fp8 is "worse." For most portraits the fp8 InfuseNet is visually indistinguishable from bf16, and it's what lets you actually finish a run on consumer hardware. Pick bf16 for the pride, pick fp8 for the workflow.
• The first load looks frozen while it downloads gigabytes. It's not hung - check your network and wait it out.