NF4 (4-bit) quantization of the HunyuanImage-3.0 Instruct Distil model (v2). The most accessible option -- fits on a single 48GB GPU with ~6x faster generation (8 steps vs 50). Best balance of speed, quality, and VRAM.

v2 uses improved quantization with more precise skip-module selection, keeping attention projections and critical embedding layers in full BF16 precision for better image quality.

• Instruct model -- supports text-to-image, image editing, multi-image fusion
• Chain-of-Thought -- built-in think_recaption mode for highest quality
• NF4 quantized -- ~48 GB on disk
• 8 diffusion steps (CFG-distilled)
• Block swap support -- offload transformer blocks to CPU for lower VRAM
• ComfyUI ready -- works with Comfy_HunyuanImage3 nodes

Component	Memory
Weight Loading	~29 GB weights
Inference (additional)	~12-20 GB inference
Total	~41-49 GB

Recommended Hardware:

• Single 48GB GPU (RTX 6000 Ada, RTX PRO 5000, A6000)
• With block swap: may work on 24GB GPUs (swapping ~20 blocks)

• Architecture: HunyuanImage-3.0 Mixture-of-Experts Diffusion Transformer
• Parameters: 80B total, 13B active per token (top-K MoE routing)
• Variant: Instruct Distil (CFG-Distilled, 8-step)
• Quantization: 4-bit NormalFloat (NF4) quantization via bitsandbytes with double quantization
• Diffusion Steps: 8
• Default Guidance Scale: 2.5
• Resolution: Up to 2048x2048
• Language: English and Chinese prompts

This is the CFG-Distilled variant:

• Only 8 diffusion steps needed (vs 50 for the full Instruct model)
• ~6x faster image generation
• No quality loss -- distilled to match the full model's output
• cfg_distilled: true means no classifier-free guidance needed

Layers quantized to NF4:

• Feed-forward networks (FFN/MLP layers)
• Expert layers in MoE architecture (64 experts per layer)
• Large linear transformations

Kept in full precision (BF16):

• VAE encoder/decoder (critical for image quality)
• Attention projection layers (q_proj, k_proj, v_proj, o_proj)
• Patch embedding layers
• Time embedding layers
• Vision model (SigLIP2)
• Final output layers

This model is designed to work with the Comfy_HunyuanImage3 custom nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Comfy_HunyuanImage3

1. Download this model to your preferred models directory
2. Use the "Hunyuan 3 Instruct Loader" node
3. Select this model folder and choose nf4 precision
4. Connect to the "Hunyuan 3 Instruct Generate" node for text-to-image
5. Or use "Hunyuan 3 Instruct Edit" for image editing
6. Or use "Hunyuan 3 Instruct Multi-Fusion" for combining multiple images

The Instruct model supports three generation modes:

Mode	Description	Speed
image	Direct text-to-image, prompt used as-is	Fastest
recaption	Model rewrites prompt into detailed description, then generates	Medium
think_recaption	CoT reasoning -> prompt enhancement -> generation (best quality)	Slowest

Block swap allows running INT8 and BF16 models on GPUs with less VRAM than the full model requires. The system keeps N transformer blocks on CPU and swaps them to GPU on demand during each diffusion step.

blocks_to_swap	VRAM Saved	Recommended For
0	0 GB	96GB+ GPU (no swap needed)
4	~5 GB	80-90GB GPU
8	~10 GB	64-80GB GPU
16	~19 GB	48-64GB GPU
-1 (auto)	varies	Let the system decide

This is a quantized derivative of Tencent's HunyuanImage-3.0 Instruct.

• License: Tencent Hunyuan Community License

• Original Model: Tencent Hunyuan Team
• Quantization: Eric Rollei
• ComfyUI Integration: Comfy_HunyuanImage3

This model inherits the license from the original Hunyuan Image 3.0 model: Tencent Hunyuan Community License