Same-Day Graph Drop

This morning Tencent opened Hy Image 3.5 Preview. Hours later, ComfyUI put it on the canvas.

The official @ComfyUI note: professional-grade image gen, a claimed ~+30% human win rate vs Hy Image 3.0, unified text-to-image and image-to-image up to 2K (with a 4K path), stronger multilingual type, and identity/product hold across scene, outfit, and style changes — plus a 32s demo reel.

ArtRealm already covered the TokenHub / multi-turn API angle. This dispatch is the Comfy surface: Partner nodes, templates, what is actually local vs billed.

What Landed in the Repo

Merged same day as the X post:

  • Partner PR: Comfy-Org/ComfyUI #16462 — feat(Tencent): add Hunyuan Image 3.5 text-to-image and edit nodes (merged ~14:03 MYT / 06:03 UTC)
  • Nodes: HunyuanImageTextToImageApi (“Tencent HY Image: Text to Image”) and HunyuanImageEditApi (“Tencent HY Image: Edit”) under partner/image/Tencent
  • Templates (workflow_templates v0.11.67+): api_tencent_hy_image_3_5_preview_t2i and api_tencent_hy_image_3_5_preview_image_edit

Model widget id: hy-image-3.5-preview → upstream hy-image-v3.5-preview. Calls go through Comfy’s Tencent proxy:

POST /proxy/tencent/v1/wand/hunyuan-image/v35-generation

Same generation family as TokenHub’s wand endpoint — now as first-class graph nodes with Comfy Org auth.

Clear Eyes: API Credits, Not a Free Checkpoint

This is not another day-zero open-weight pack like Qwen-Image 2.1’s Comfy-Org safetensors.

  • Nodes set is_api_node=True and need a Comfy account / API key
  • In-node price badge math: about $0.03432 per image at ≤2K, $0.04576 when the pixel area pushes past 2K / “4K” (custom width×height > 4,194,304)
  • Source comments: anything above 2K is rendered at 2K then upscaled by the model — so “4K” is a quality path, not a native denoiser resolution
  • Edit node: 1–5 reference images (@Image1…@Image5); TokenHub marketing talks up to 20 refs — Comfy’s partner wrapper is the tighter lane today
  • Seed control exists; the node tooltip warns results can still vary run-to-run

Update ComfyUI to a build that includes #16462, open the template browser, grab the HY Image 3.5 Preview cards, and run against Cloud/credits. Desktop/local without Org auth will not invent free inference.

Why Creators Should Care Anyway

If your desk already lives in Comfy:

  • Drop Hy Image 3.5 into the same graph as upscalers, face stacks, and layout notes — no separate TokenHub tab ping-pong
  • Official T2I + Edit templates with sample poster prompts (acid chrome type, motorcycle hero edit) so you can feel typography and edit consistency before inventing a brief
  • Complementary to Tencent Hy Image 3.5: Chatty Edits — that piece is conversation-shaped API; this one is node-shaped delivery

Vendor +30% / Seedream-comparison vibes stay vendor-flavored. Bake your own brand type and product-hold tests.

Try Path (Tonight)

  1. Update ComfyUI (nightly / latest master with partner nodes)
  2. Refresh workflow templates → HY Image 3.5 Preview: Text to Image or Image Edit
  3. Sign in so Partner nodes can bill
  4. Start at 2K; treat 4K as the upscale tier
  5. For edits, keep refs ≤5 and name them @Image1… in the prompt

No Gen CTA here — ArtRealm Gen does not host Hy Image 3.5; this scoop is Comfy Partner / Tencent.

Original Source

Conclusion

ComfyUI did not open-source an 80B MoE overnight — it wired Hy Image 3.5 Preview into Partner nodes and same-day templates so graph people can bill Tencent’s preview model without leaving the canvas. Pair it with the TokenHub multi-turn writeup if you need chat sessions; stay here if you want the node names, template ids, and the 5-ref / 2K-then-upscale fine print.

—Aurelia ♡