I’ve been working on native ComfyUI support for Tencent’s HunyuanImage 3.0, the 80B mixture-of-experts image model (13B active per step). It isn’t a wrapper around Tencent’s pipeline: it uses the normal KSampler, the normal VAE Decode and ComfyUI’s own memory management, which streams the experts from system RAM so the model fits on one consumer GPU.
What’s in the gallery (all Instruct-Distil, 8 steps):
- 4-bit vs int8, same prompt and seed. Times are the whole generation on an RTX 3090.
- image editing with the 4-bit weights. The instruction is at the top of each image.
- style transfer with the 4-bit weights: two input images, the photo and a style reference.
These are picked from a bigger run: 60 prompts × 2 formats, 60 edits and 14 styles, one seed each, no rerolls. Most of the edits worked; a few didn’t (snow that barely shows, a logo it wouldn’t remove, a “make it night” that stayed day). The text-to-image prompts come from popular prompt posts on X.
What you get
- All three models: Instruct-Distil (8 steps, the one to start with), Instruct (50 steps) and Base
- Text-to-image, image editing, and multi-image fusion with up to 3 input images (that’s how the style transfer works)
- Optional prompt rewriting: the model expands your prompt first (slow, about 1 s per token)
- Optional Spectrum speed-up: about 3.4× faster for the 50-step models
- Ready-made weights: 4-bit W4A8 (44 GB), int8 (76 GB) and bf16 (150 GB, mostly for comparisons)
- Example workflows for each model and task
- Nothing to
pip install
Speed (Instruct-Distil, about 1 megapixel):
| GPU | 4-bit W4A8 | int8 |
|---|
| RTX 4090 | ~22–26 s | ~47–49 s |
| RTX 3090 | ~29 s | ~54 s |
It also runs with only 16 GB or 12 GB of VRAM (~30 s and ~32 s per image on a 4090 limited to that).
What you need
- An NVIDIA GPU with 12 GB+
- Lots of system RAM: ComfyUI held about 50 GB with the 4-bit file loaded. This is the real requirement, since the experts live in RAM and stream over PCIe every step.
- A recent ComfyUI (late September 2026 or newer)
Links
Happy to answer questions. If something breaks, open an issue on GitHub with the traceback.
