TensorScale Hunyuan Image 3 Fast (Soze)

Most image nodes generate or edit. This one can do four things, because Tencent's Hunyuan Image 3 has a bot_task mode that wraps a reasoning step around generation: image, recaption, think, and think_recaption. Plain image is just text-to-image (or edit-from-references). Turn on think and the model reasons about your prompt before rendering - good for complex scenes where the first draft otherwise comes back literal and dumb. recaption generates a caption from a reference image; think_recaption does both. It's the rare hosted image node with a thinking step, and it's served here through TensorScale.

The inputs

Required: prompt, bot_task, and seed. The prompt does double duty - "what to generate, or the edit to apply to the references." So for a pure generate: "a brass octopus in a steampunk apothecary." For an edit, connect references and prompt the change: "turn the sky behind the lighthouse into a storm."

• bot_task - the four-way dropdown above. Default image. Start there; think_recaption is a fun party trick but slower and pricier.
• seed - defaults to 43. Set it to lock a take.

The reference path is deliberately capped: image_1 and image_2 only - the API's max is two. That's the whole reference surface, so this node is for single-edit or paired-reference jobs, not collage. image_url_1/image_url_2 override the sockets with public HTTPS URLs. use_fal_upload (default off) pushes connected images to FAL's CDN instead of inlining them as base64 - flip it if the API complains about body size. image_format chooses PNG or JPEG for the inline encoding.

How it works

TensorScale's /v2/hunyuan-image-3-instruct/fast runs synchronously with a model-scoped key (TENSORSCALE_API_KEY_HUNYUAN_IMAGE_3, falling back to TENSORSCALE_API_KEY). It's a fast tier, so the default timeout is 600s rather than the 1800s the video nodes use - but it's still one POST with the socket open until the image streams back. The result is saved to disk and returned as a standard IMAGE tensor.

Outputs: image (wire it to a preview or the next node), image_path, request_id, config (run summary), and status.

Install and key

From the Soze pack ("Quality of Life Nodes for ComfyUI"), via ComfyUI Manager or:

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

Restart, then set TENSORSCALE_API_KEY or TENSORSCALE_API_KEY_HUNYUAN_IMAGE_3 (environment variable or .env in the repo root). Don't commit keys.

Common issues

• Third reference ignored. The API caps at two - image_3 doesn't exist here. Plan the edit for two inputs max.
• think tasks feel slow. The reasoning step is extra work on top of generation; the longer wait is the model thinking, not a hang. Drop to image for drafts.
• Body-size errors with a large reference. Turn on use_fal_upload (needs FAL_KEY) or use the image_url widgets with hosted images.