Black Forest Labs on X

The 21 people who praise it talk about output, not price or features: realistic lighting, clean image editing, video edits, one clip built from text-to-video, video-to-video and image-to-video together, an Arabic prompt that rendered connected script, and a film teaser made on the model.

9 people treat the first places as the point — #1 on Physics-IQ across fluid dynamics, optics and mechanics, tied for first on image-to-video — and name Veo, Sora and Seedance as what it beat. 7 people use numbers for speed: a 10-second clip in about 12 seconds, 2.0 seconds per 1024² image on an RTX 5090, 6.3 seconds on a 16 GB card, and a "speed to quality ratio" praised for creators.

The 8 disappointed people are blunt: "censored trash", a model that failed on every try, and doubts that the rankings mean anything — one asks how much the rank cost. Some say FLUX 2 Klein 9B kept faces better when editing people. 4 people like it with a hedge, among them "close enough" at 1080p and one whose only complaint is crushed blacks.

The 4 people unhappy with the demos are asking for evidence they can judge. They name a 720p example video, a product name in an announcement that points at something that "no longer exists", upscaling demos that are "just a blurry video with a white line sweeping left and right", and before/after shots cut together too fast to see the improvement.

Filters come next: 3 people say the models are heavily censored, one calling them "the strictest filters in the industry" and saying nothing would generate; another reads "upscale any video" as "just what passes your purity tests". 3 people prefer earlier models for their own work — the older Klein keeping faces and editing people better than the newest one, with one saying the earlier model works best in their own cases.

Behind these, 2 people are held back by missing open weights and a permissive licence, one refusing to use it before "MIT or Apache 2.0". 2 people doubt the rankings, 2 find generation slow — the slowest median in a night's test, and a ten-second clip taking twelve seconds — and 2 report output they cannot depend on, from generations that fail every time to the same seed giving different bytes on different machines.

The Topaz talk is competitive: one says being twice as fast as Topaz "is gonna hurt some renewals", another that Topaz Starlight is still the best upscaler, and two ask what kind of upscaler BFL shipped — a precision type like Topaz Starlight Precise, a creative one, or Topaz under another name.

Qwen-Image-2.1 gets 2 people. One says Qwen excelled more than the newest Klein at editing images with people in them; a benchmark run across four machines found Qwen winning three of four motifs on a Mac Studio but taking far longer per image, 29.9 s against 8.8 s, so the pick still went to FLUX on speed.

One person's multi-post benchmark series carries the MOSS-Image-SFT comparison: MOSS's images win every motif on a 16 GB card but take 40.9 s against 6.3 s, and speed "matters in daily use", so FLUX wins the pick. Two single posts round it out — one prefers FLUX's speed to a further quality bump over Seedance 2.5, another asks how it compares with Ideogram 4.5 — and one post treats Odyssey 3 Pro as a head-to-head race on the Physics-IQ image-to-video board.

The open-weights questions are short and impatient — "will you guys release the open weight?", "WHEN OPEN WEIGHTS?" — and they extend to the video model's weights, which one person says should be easy to answer since some companies already have them for post-training. One asks about a release because the weights "might" come.

The upscaler questions are practical: the 20-second input cap against four-minute clips edited in DaVinci Resolve, whether audio passes through untouched or gets re-encoded on the way to 4K, and whether "upscale any video" covers any video at all. 2 people ask what kind of upscaler it is — precision like Topaz Starlight or BFL's own.

2 people ask when a Dev release arrives rather than Klein, and 2 ask whether the video is convincing beyond speed: whether it does dialogue and how motion consistency holds up. One asks what the price will be after the promotional period ends, and one wants to know whether changing a small detail keeps the rest of the image stable.

The Dev request is the only one made twice, both as an impatient ask: "we want Image 3 Dev!!! not klien!!! when we are having that?" and a plain "Waiting on Flux 3 Dev". One of those posts is filed as a question too, so the same people appear in both sections.

The rest are single asks. One user is still waiting for a proper web UI for FLUX, does not like the Playground, and wants something like Ideogram's interface. One, praising the model's video, says the next update needs reference audio and reference video support — the only wish that comes attached to praise rather than a complaint.

The last wants to see a small detail changed without the rest of the image shifting, which matches a question asked elsewhere on the page about whether a small edit stays stable. None of these people say what they would pay or when they need it.