Black Forest Labs — the frontier AI analysis lab growing visible generative AI fashions — at this time launched the FLUX.2 household of state-of-the-art picture technology fashions.
FLUX.2 is filled with new instruments and capabilities, together with a multi-reference characteristic that may generate dozens of comparable picture variations, in photorealistic element and with cleaner fonts — even at scale.
NVIDIA has labored with Black Forest Labs and ComfyUI to make the fashions accessible with FP8 quantizations and RTX GPU efficiency optimizations at launch, reducing the VRAM required to run them by 40% and enhancing efficiency by 40%.
Requiring no particular software program bundle to run, the fashions can be found immediately in ComfyUI.
State-of-the-Artwork Visible Intelligence
Photos generated by FLUX.2 are photorealistic, even at scale, that includes as much as 4 megapixel decision with real-world lighting and physics to eradicate that “AI look” that undermines visible constancy.
The fashions add direct pose management to explicitly specify the pose of a topic or character in a picture, in addition to ship clear, readable textual content throughout infographics, consumer interface screens and even multilingual content material. Plus, the brand new multi-reference characteristic allows artists to pick as much as six reference photos the place the fashion or topic stays constant — eliminating the necessity for intensive mannequin fine-tuning.
For an entire overview of recent FLUX.2 options, learn Black Forest Labs’ weblog.
Optimized for RTX
The brand new FLUX.2 fashions are spectacular, but in addition fairly demanding. They run a staggering 32-billion-parameter mannequin requiring 90GB VRAM to load utterly. Even utilizing lowVRAM mode — a well-liked setting that permits artists to solely load the energetic mannequin at a time — the VRAM requirement continues to be 64GB, which places the mannequin nearly out of attain for any shopper card to make use of successfully.
To broaden FLUX.2 mannequin accessibility, NVIDIA and Black Forest Labs collaborated to quantize the mannequin to FP8 — lowering the VRAM necessities by 40% at comparable high quality.
And to make this mannequin accessible on GeForce RTX GPUs, NVIDIA has partnered with ComfyUI — a well-liked utility to run visible generative AI fashions on PC — to enhance the app’s RAM offload characteristic, generally known as weight streaming.
Utilizing the upgraded characteristic, customers can offload elements of the mannequin to system reminiscence, extending the accessible reminiscence on their GPUs — albeit with some efficiency loss, as system reminiscence is slower than GPU reminiscence.
NVIDIA has additionally been collaborating with ComfyUI to optimize mannequin efficiency on NVIDIA and GeForce RTX GPUs, together with optimizations for FP8 checkpoints.
Get began with FLUX.2 at this time. Replace ComfyUI and take a look at the FLUX.2 templates, or go to Black Forest Labs’ Hugging Face web page to obtain the mannequin weights.
Plug in to NVIDIA AI PC on Fb, Instagram, TikTok and X — and keep knowledgeable by subscribing to the RTX AI PC publication.
