IronLeo78-FaceFluxCapacitor
Swap a head easily and FAST, in one node: load a body and a head, press Run. Custom tuned checkpoint, optional LLM, preview with Save button, no wires.
IronLeo78-FaceFluxCapacitor
Swap a head easily and FAST, in one node. Load a body, load a head, press Run.
USE NODES 2.0!
Using the original (classic) nodes is discouraged: this node must be used with Nodes 2.0. It is WORTH it: the picture block, the buttons and the theme are built for it. When an IronLeo78-FaceFluxCapacitor node opens and Nodes 2.0 is off, the pack switches it on for you and shows a notice (if the node looks odd, press Ctrl+F5). To keep the classic canvas, run
localStorage.setItem("ironleo78_keep_classic", "1")in the browser console: you then only get the notice.
Download the checkpoint first
https://huggingface.co/Ironleo78/IronLeo78-FaceFluxCapacitor
The checkpoints are always published there. Put
IronLeo78-FaceFluxCapacitorLight.Checkpoint.safetensors(LOW, the default, required) inComfyUI/models/checkpoints.IronLeo78-FaceFluxCapacitor.Checkpoint.safetensors(HIGH) is optional: you only need it if you switch to HIGH.
Models
| File | Folder | Download |
|---|---|---|
| IronLeo78-FaceFluxCapacitorLight.Checkpoint.safetensors (LOW, default, required: Klein 4B + Qwen3 4B) | models/checkpoints | Hugging Face |
| IronLeo78-FaceFluxCapacitor.Checkpoint.safetensors (HIGH, optional: Klein 9B + Qwen3 8B) | models/checkpoints | Hugging Face |
| Huihui-Qwen3-VL-4B-Instruct-abliterated-int8_convrot.safetensors (optional: only needed if you switch LLM ON, which is off by default; the same small text encoder as IronLeo78-MTV) | models/text_encoders | Hugging Face |
If a file you use is missing, the node stops with a message that names it and where to get it. Only the LOW checkpoint is required.
Install
Put this folder (ironleo78_facefluxcapacitor) in ComfyUI/custom_nodes/, restart ComfyUI, press Ctrl+F5. The node is under IronLeo78. Use it with Nodes 2.0 (the pack switches it on by itself when the node opens).
It runs even on a very old PC: tested on a GTX 1080 (8 GB VRAM), i7-8700K, 16 GB RAM, with MODEL LOW and LLM OFF (about 4.5 minutes for the first run).
Use
- BODY (Image 1, left): a medium shot or wider (shown on LOW only, the light model needs it). The image that is kept. Lighting, environment, background, clothes, everything stays.
- HEAD (Image 2, right): a big face, close-up, filling most of the picture (shown on LOW only, the light model needs it). The head that is swapped in (hair, eye color, nose structure).
- Click a picture box or drop a file on it.
- RESULT: a preview. Nothing is written to your output folder until you press Save.
- LLM OFF / ON: OFF uses the tuned prompt as is. ON lets the small LLM read the BODY picture and describe its facial expression, so the new head copies it. Both work.
- MODEL LOW / HIGH: LOW (pink, default) uses the light 4B checkpoint, which needs much less memory. HIGH (green) uses the full 8B checkpoint. Only one is kept in memory.
- prompt: your own words, added to the end of the tuned prompt. The tuned prompt itself is hardcoded and cannot be broken from here.
The node needs no wires: everything is inside it. It has one optional image output (the swapped picture) for later use, for example to feed another node. Everything else (4 steps, CFG 1, lcm sampler, 1 megapixel, negative prompt) is hardcoded and tuned.
Saved files go to output/IronLeo78-FaceFluxCapacitor_00001_.png and keep the workflow in their metadata.
Tests
python -m pytest tests
License
MIT