HunyuanImage 3.0 in ComfyUI — compare every image

Three models, three weight formats, Spectrum off and on, five prompts. Every image is 1024×1024, rendered with the same seed per prompt at each model's recommended settings on one RTX 4090, through the ComfyUI-HunyuanImage3 nodes.

⇆

Drag across the image to move the divider.

Every version of this prompt

Prompt rewriting

The same prompt as typed, with rewrite only (the model rewrites the prompt, then draws), and with think + rewrite (it reasons about the request first). Instruct-Distil, W4A8, same image seed. The rewriting runs at about 1 s per token on the GPU, before the 8 image steps.

Image editing

The Instruct-Distil portrait above, edited by instruction (Instruct-Distil, W4A8, 8 steps). The last two repeat the setting edit with prompt rewriting, which here looks at the input image.