Qwen-Image-2.0: Professional infographics, exquisite photorealism

Original Article ↗ Hacker News Discussion ↗

Visual realism, artifacts, and uncanny feel

  • Many commenters find the “photorealistic” samples subtly wrong: over‑crisp textures, flat/HDR‑like lighting, weak or inconsistent shadows, and an overall “weightless” look.
  • Depth of field is a recurring tell: sometimes absent, sometimes present but physically incorrect (blur amount vs distance/zoom), making scenes feel composited or “focus‑stacked.”
  • One long technical comment attributes this to diffusion models learning archetypal “texture brushes”: everything gets rendered in perfect focus at fixed scales, so surfaces have too much visible detail at distance (a “doll clothes” / video‑game‑character effect).
  • Some report physical discomfort or mild nausea staring at the images, attributing it to subtle violations of real‑world cues.

Comparisons to other image models

  • Opinions vary on how Qwen-Image-2 stacks up against Midjourney, Flux, Z‑Image, GPT Image, and “Nano Banana Pro” (Gemini‑based).
  • Debate centers on three competing goals: photorealism, aesthetics, and prompt adherence.
  • Midjourney is seen as still unmatched for aesthetics but weak on prompt following and editing; some say modern local models like Flux/Z‑Image/Qwen are catching up or surpassing it overall.
  • Others argue some SOTA models now drift into an “AI slop” aesthetic when pushed toward stronger prompt alignment.

Prompt following, infographics, and reliability

  • The model’s complex prompt following and editing are widely praised, especially for detailed scenes and multi-step workflows.
  • Infographics are seen as a mixed bag: technically impressive layouts but often “cognitive slurry” that doesn’t actually clarify information—though this is blamed partly on users’ poor design skills.
  • A comic‑panel example from the blog reproduces perfectly when re‑prompted verbatim, but small prompt changes cause layout breakdowns (wrong grid sizes, missing panels, language switching to Chinese), raising questions about robustness.

The “horse riding man” image and controversy

  • The horse‑standing‑on‑a‑man image draws strong reactions: disturbing, “revenge porn,” or darkly comic; many think it’s an odd choice for a flagship demo.
  • A translated internal prompt shows extremely detailed specification of the scene, including that the man is white and “subdued.” This undermines the notion it was an accidental interpretation of “horse riding a man.”
  • Some argue it’s a deliberate “horse versus human” benchmark (like prior “horse rides astronaut” tests); others see it as a year‑of‑the‑horse visual metaphor (“East trampling West”) or tie it to Chinese memes and historical statuary.
  • Several call it tone‑deaf given global racial/political context, especially since most other examples feature East Asian faces while the one humiliating image uses a white man.

Openness, censorship, and Chinese context

  • Weights are not yet released; based on Qwen’s history, some expect open weights in weeks, others are skeptical and accuse the blog of “coming soon” open‑washing.
  • It’s noted that previous Qwen image weights did ship under Apache‑style licenses, and that Alibaba generally doesn’t pretend its largest models are open.
  • A user test prompt about Tiananmen and “Tank Man” gets blocked with a “content security” warning, suggesting strong server‑side censorship. It’s unclear whether this is encoded in the model or just a service‑level filter.
  • A commenter with China experience says local sentiment is largely enthusiastic about AI, seen as opportunity and status, with less anti‑AI backlash than in the West (though some hostility exists).

Technical notes on Qwen-Image-2 and ecosystem dynamics

  • Qwen‑Image‑2 is described as a 7B unified image+edit model, down from the prior ~19B model that required large GPUs and had known high‑frequency VAE artifacts and odd timestep embeddings.
  • It now uses the newer Qwen‑3‑VL backbone; some expect better quality and accessibility on modest hardware, aligning it with Z‑Image Turbo and Flux.2 Klein in the “post‑SDXL” local SOTA race.
  • Vertical Chinese typography in the demos is called out as slightly off (wrong punctuation forms).
  • Some earlier Qwen‑Image‑2512 users report poor English text rendering and spelling, inconsistent with the new blog samples.

Local tooling and commoditization

  • For Linux/local use, people recommend ComfyUI (despite a steep initial learning curve), Stable‑diffusion.cpp, Koboldcpp with image support, Lemonade (for AMD), and various “manager” tools.
  • There’s broad agreement that image models are rapidly commoditizing: SOTA shifts every few months, while many creators remain productive with older SDXL‑era models.
  • Several argue the main bottleneck is now the human “director” (prompting, iteration, and taste), not the raw model capabilities.