ComfyUI and open-source status
Hy Image 3.5 in ComfyUI and its open-source status
You can use Hy Image 3.5 Preview inside ComfyUI today, but not the way you used HunyuanImage 3.0. There is no checkpoint to download. The official route is a pair of partner nodes that call Tencent's servers.
The short answer
- ComfyUI: yes, through the partner nodes "Tencent HY Image: Text to Image" and "Tencent HY Image: Edit". They run remotely, so no model download or local GPU is needed.
- Open weights: no. Tencent has not released weights for Hy Image 3.5 Preview, and has not announced that it will.
- Hugging Face, GGUF, LoRA: nothing official exists for 3.5. The open repositories are for HunyuanImage 3.0 and its Instruct variants.
How the ComfyUI partner nodes work
ComfyUI documents two nodes under its Tencent partner category. "Tencent HY Image: Text to Image" generates from a prompt, with optional prompt rewriting. "Tencent HY Image: Edit" edits or combines up to five reference images from a written instruction. Both send the job to Tencent through the Comfy API and return the image to your graph.
| Setting | What the ComfyUI docs list |
|---|---|
| Resolution | 1K, 2K, 4K, or custom dimensions in multiples of 16 up to about 4096x4096 |
| 4K option | Described as rendered at 2K and upscaled |
| Aspect ratio | Auto or manual; auto is not available at 4K |
| Reference images (Edit node) | Up to 5 |
| Reference detail | Standard (1024x1024) or high (2048x2048) |
| Watermark | Optional, off by default |
| Templates | api_tencent_hy_image_3_5_preview_t2i and api_tencent_hy_image_3_5_preview_image_edit |
To use them you need a current ComfyUI build (the docs ask local users for the nightly version), to be signed in to a Comfy account, and a network that can reach the Comfy API. Partner nodes are paid per run through that account; the tutorial does not state a price, so read the cost shown in ComfyUI before queueing a batch.
Note the gap with the raw API. Tencent's API guide documents up to 20 reference images and native output at an explicit 4096x4096; the node exposes five references. If you need the higher limits, call the API directly.
Is Hy Image 3.5 open source?
No. Hy Image 3.5 Preview was released on September 22, 2026 as a hosted model. Tencent published an API, product integrations and a price, but no weights, no parameter count, no technical report and no statement about a future open release. That means there is nothing to load with a checkpoint loader, no GGUF quantisation to run on a smaller card, and no base to train a LoRA on.
If a download claims to be Hy Image 3.5 or HunyuanImage 3.5 weights, treat it with suspicion. As of this review Tencent's official repositories stop at 3.0.
What you can run locally: HunyuanImage 3.0
The previous generation is open. HunyuanImage 3.0 was released on September 28, 2025 with weights on Hugging Face and code on GitHub under the Tencent Hunyuan community licence. Its model card describes a unified autoregressive multimodal model, built as a mixture of experts with 64 experts, 80 billion parameters in total and about 13 billion active per token. On January 26, 2026 Tencent added HunyuanImage 3.0 Instruct, which brings image-to-image editing and multi-image fusion, and Instruct-Distil, an 8-step checkpoint for cheaper inference.
That model is large. Community ComfyUI node packs exist for it and rely on quantised variants and offloading to fit consumer hardware; read each project's own requirements before you plan around one.
| Hy Image 3.5 Preview | HunyuanImage 3.0 family | |
|---|---|---|
| Weights | Not released | Open on Hugging Face and GitHub |
| ComfyUI | Official partner nodes, remote | Community custom nodes, local |
| Hardware | None; runs on Tencent's servers | A large GPU, or quantised builds |
| Cost model | Per image | Your own compute |
Hy Image 3.5 without ComfyUI
If you want the Hunyuan image line without setting up either ComfyUI route, this workspace runs Hy Image 3.5 Preview by default, plus Hunyuan Image 3.0 and Hunyuan Image 3.0 Instruct, as hosted models. Write a prompt, pick the model, and the credit cost is shown before you submit. The Hy Image 3.5 edit mode accepts up to 20 reference images here. New accounts receive one free Hy Image 3.5 Preview image.
On hyimage35.pro, Hy Image 3.5 Preview is the default model, with 2K or 4K output and an edit mode that takes up to 20 reference images. Hunyuan Image 3.0 and the rest of the image catalog are listed under their own names. This site is independent and not affiliated with Tencent.
Hy Image 3.5 in ComfyUI FAQ
Can I download Hy Image 3.5 for ComfyUI?
No. There is no Hy Image 3.5 checkpoint. ComfyUI support comes from the Tencent HY Image partner nodes, which send your request to Tencent's servers and return the result.
Is there a Hy Image 3.5 GGUF or quantised version?
No. GGUF and other quantised builds require released weights, and Tencent has not released any for 3.5. The quantised community builds you may find are for HunyuanImage 3.0.
Do the ComfyUI nodes need a GPU?
No. According to the ComfyUI docs both nodes run on Tencent's servers through the Comfy API, so no model download or local GPU is required. You do need a Comfy account and credits.
Why does ComfyUI allow 5 reference images when the API allows 20?
They are different surfaces. Tencent's API guide documents up to 20 references per request, while the ComfyUI Edit node and Tencent's launch materials describe five. Use the API directly if you need more than five.
Can I train a LoRA for Hy Image 3.5?
Not at present. LoRA training needs access to the model weights, which are not public for 3.5. HunyuanImage 3.0 is the open model in this family.
Will Hy Image 3.5 be open-sourced later?
Tencent has not said. HunyuanImage 3.0 was released with open weights, but no equivalent announcement has been made for 3.5, and this page does not predict one.