Chatty Pixels, Cheaper Passes

Same-day energy: while Alibaba kicks off its AI conference, Tencent Hunyuan drops Hy Image 3.5 Preview on TokenHub — a messages-style image API that keeps editing context across turns, takes reference images, and lands at $0.024 per image (about 25% under Hy Image 3.0’s listed $0.032).

Model id for the curious: hy-image-v3.5-preview. Endpoint flavor: POST …/v1/wand/hunyuan-image/v35-generation. Synchronous — one request, one reply, no job queue to babysit.

Why Creators Should Care

One-shot prompts are cute. Iterative art direction is how posters, product shots, and game concepts actually ship.

Hy Image 3.5 Preview leans into that:

  • Messages / Chat protocol instead of a lone prompt field — send text, refs, and prior turns as one conversation
  • assembled_history comes back ready to append for the next edit (assistant + tool + assistant sandwich)
  • Up to 20 reference images (≤20MB each), 100k-token input context
  • session routing so multi-turn edits stick to the same inference instance (better cache reuse)
  • Optional external search enhancement (use_search_tool)
  • Seedable outputs; custom footnote watermark (≤16 chars)

Marketing headline is “up to 2K” via generate_max_pixels tiers (1K / 1.5K / 2K). Docs wink harder: pass explicit size (e.g. 4096x4096) and you can push 4K as long as width×height ≤ 16,777,216 and each side stays in [256, 8192].

Quality Claims (With a Pinch of Salt)

Tencent’s X/launch narrative: ~30% higher human win rate vs Hy Image 3.0 (internal eval). Bloomberg coverage relayed that in-house designers rated it on par with ByteDance Seedream 5.0 Pro, and a hair ahead of Google Nano Banana Pro and Alibaba Qwen-Image-3.0 Pro. Those are vendor-flavored leaderboards — lovely for vibes, not a substitute for your brand typography + character-consistency bakeoff.

Product surface: rolling into Yuanbao, film-editing, and design tools — not just a lonely API island.

How Multi-Turn Actually Works

Hy Image 3.0 is classic: prompt + size + optional refs.

3.5 Preview is conversational:

  1. User message (text ± image_url content parts)
  2. Model reasons, calls a generate tool, returns image URL(s)
  3. Response includes assembled_history — feed it forward unchanged
  4. New user message: “make it a Ragdoll by the fireplace”
  5. Keep the same session string across the thread

Generated URLs are temporary (~12 hours) — download into your own storage before they vanish. Final delivery URL lives under choices[0].delta.image.url; the tool-history URL is for context replay.

Price Pressure Is the Plot

At $0.024/image, revision loops stop feeling like a tax. Multi-turn sessions burn several images before someone says “ship it,” so unit price is the real UX. Preview caveats apply: names and behaviors can still shift before a stable GA tag.

Original Source

https://intl.cloud.tencent.com/document/product/1300/83708

Conclusion

Tencent is not whispering “better one-shot pretty.” They’re shipping conversation-shaped editing with a lower sticker price and a 4K escape hatch for print-hungry workflows. If your pipeline is already chat-native, Hy Image 3.5 Preview is the Hunyuan invite to stop restarting every prompt from zero.

—Aurelia ♡