Hunyuan Instruct Image Edit

"Hunyuan Instruct Image Edit" is the node that does what it says on the tin: hand it an image and an instruction like "change the background to a sunset scene," and it edits the image. No masks, no inpaint gymnastics, no ControlNet - just a sentence. It runs on the Instruct family of HunyuanImage-3.0, which is autoregressive, so the model literally reasons about your instruction before it generates. That's the whole differentiator: you're not describing pixels, you're having a conversation with an image model.

The inputs that matter

• image - the IMAGE to edit (from Load Image).
• instruction - your edit, in plain language. The model understands English and Chinese.
• bot_task - the biggest decision. image applies your instruction as-is (fast). recaption has the model rewrite your instruction into a detailed edit description first (better for vague asks). think_recaption adds a chain-of-thought pass that reasons about what to change and what to preserve - the tooltip calls it "BEST QUALITY," and the warning is equally blunt: the CoT text generation adds minutes before editing even starts. Use image when your instruction is already precise.
• seed - -1 for random.
• resolution - auto (default) with align_output_size: true matches the input image's dimensions (snapped to multiples of 16), which keeps 2K+ inputs at 2K+. Set a fixed preset to force an output size instead.
• steps / guidance_scale - both default to -1 (auto: 8 steps for Distil, 50 for full; CFG ~2.5). Push steps to 60–80 at 2K+ to reduce flow-matching artifacts.
• flow_shift - the denoising schedule shape, default 2.8. 2.0–2.5 for portraits (sharper), 3.5–5.0 for landscapes/illustrations.
• max_new_tokens - cap for the CoT/rewrite output, default 2048.
• system_prompt - dynamic (default) picks the right prompt for your bot_task; en_unified covers every edit mode including style transfer and text editing.

Outputs: image (the edit), cot_reasoning (a STRING of the model's thinking - fascinating to read, and genuinely useful for debugging why an edit went sideways), and status.

Workflow shape

Load Image → Hunyuan Instruct Image Edit → Save Image

The model input comes from the Hunyuan Instruct Loader. Both must be Instruct - the base loaders are a different output type and won't plug in.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Comfy_HunyuanImage3
cd Comfy_HunyuanImage3
pip install -r requirements.txt

Restart ComfyUI, or use ComfyUI Manager (search "Comfy_HunyuanImage3"). You need an Instruct model in ComfyUI/models/ - the Distil-NF4 (~45GB) is the entry point for 48GB cards; Distil-INT8 (~81GB) is the 96GB recommendation. Since this is an autoregressive model rather than diffusion, don't expect classic img2img semantics - edits can be surprisingly strong, which is why think_recaption exists to keep the model from "improving" things you wanted preserved.

The honest caveat

It's slow. The model is 80B parameters, and recaption/think_recaption run a full text-generation pass before a single diffusion step. Budget several minutes per edit on the Distil variants, more on full Instruct. And the author's known-issues list still shows edge OOM handling on Image Edit - if a big 2K+ input blows up on a 48GB card, drop to the ~1MP preset or a smaller input rather than fighting it.