FLUX is fast and it's open source

Original Article ↗ Hacker News Discussion ↗

A new generative image model called FLUX is drawing attention for its speed and image quality, including fast inference via a new synchronous API and gains from quantization, though some users question how much of this is true model improvement versus faster delivery of results. Commenters highlight that only the distilled FLUX.1 [schnell] variant is genuinely open source while higher-quality versions are non‑commercial, prompting calls to back community efforts like OpenFLUX that remove these constraints. Alongside praise for prompt adherence and realism, people raise concerns about training data opacity, the model’s tendency toward certain photographic styles (including shallow depth of field and phone‑camera aesthetics), and the economic impact on human artists, while also exploring more modular, controllable workflows for AI‑generated art.

Name reuse and positioning

  • Many note that “Flux” is already heavily used across tech (frameworks, scripting languages, AI tools, hardware, podcasts, etc.).
  • Debate whether name-collision complaints are interesting or just noise; some argue that the extreme frequency of collisions here makes it noteworthy.

Performance claims and quantization

  • Confusion over the claim that a new synchronous HTTP API “makes models faster”; clarification that it primarily removes an extra file-fetch round trip.
  • Some feel that’s not “model faster” but “delivery faster”; post author later adds clarifying text.
  • FP16→FP8 quantization shows ~2× speedup with some quality loss; people question what product use cases justify only ~2× when “realtime” offerings are much faster.

Image quality, style, and depth of field

  • Flux is praised for quality and prompt adherence, especially for locally hosted generative systems.
  • Complaints that images often have exaggerated shallow depth of field that’s hard to remove.
  • Long back-and-forth on depth of field: is it desired artistic choice vs. outdated “sensor limitation” that AI need not reproduce.
  • Some say Flux, like Midjourney, has a recognizable “signature look.”

Architecture ideas and modular workflows

  • Several propose modular pipelines: text → scene graph → semantic segmentation → final rendering, to improve editability and composability.
  • Others respond this kind of hand-engineered decomposition has historically underperformed end-to-end learning (“bitter lesson” discussion).
  • Counterpoint: modular, editable representations may be worth some loss in raw optimality for certain workflows; tools like ComfyUI partly enable this today.

Ethics, artists, and practical uses

  • Some use Flux for blog/Substack illustrations and say they would never have paid an artist anyway; they view this as analogous to open‑source/public domain access.
  • Others argue that widespread use by blogs and media erodes the market where illustrators previously were paid, more akin to piracy economics.
  • Further nuance: high-end character/visual design is seen as harder to replace than generic illustration, and AI quality is not yet sufficient for many professional needs.

Open source vs. non-commercial

  • Only FLUX.1 [schnell] is Apache 2.0; FLUX.1 [dev]/pro are non-commercial.
  • Discussion clarifies “open source” as defined by OSI/FSF (right to use, modify, redistribute), vs. merely “source available” or inspectable.
  • Some call labeling non-commercial models “open source” misleading, since it blocks others from continuing development commercially if the originator stops.
  • OpenFLUX.1 is cited as an Apache-licensed finetune aiming to undo some distillation constraints.

Training data and privacy concerns

  • Users notice that prompts resembling camera filenames (e.g., IMG_0001.JPG + a word) yield hyper-realistic, phone-photo-like images: messy apartments, food, candid people.
  • This feels to some like peeking into private photo streams; they suspect training on social media or cloud photo stores but note there is no disclosed dataset list.
  • Others point out similar behavior in Stable Diffusion and share filename conventions that models likely picked up during training.
  • Overall: strong unease, and lack of clear information on Flux 1.1’s training data is flagged as problematic and “unclear.”

Ecosystem, access, and comparisons

  • Some users cancel Midjourney, feeling Flux and other local/open models have caught up or surpassed it; others say Midjourney’s default “look” can be changed and that Flux has its own look.
  • Pollinations and other services expose Flux.schnell via simple URLs, with claims of high throughput on a small GPU cluster; others note that “only three L40S” is still expensive for individuals.
  • A few mention alternative fast systems (e.g., Krea) and community efforts to make Flux easier to run and tune (ComfyUI, OpenFLUX).

Clarity gaps and limitations

  • Multiple commenters say the original blog post doesn’t clearly explain what Flux actually is or does for readers unfamiliar with it.
  • Hands and fine details are still often rendered poorly, indicating remaining quality limitations.
  • Questions about performance on local hardware (e.g., M1 Mac, ComfyUI setups) receive no concrete answers in the thread.