RELEASE · MODELS · #2050
Black Forest Labs releases FLUX.3 Image Klein with bounding-box layout control in Diffusers
Black Forest Labs published FLUX.3 Image Klein to the Hugging Face Hub and ships open checkpoints that run in Hugging Face Diffusers. The model supports text-to-image and multi-reference image-to-image generation, adds bounding-box layout control for region placement and image edits, includes a distilled checkpoint (1024×1024 ~2s on an H100, 16 steps by default), quantized single-file transformer weights (bf16, fp8r, nvfp4), and LoRA/DreamBooth training scripts.
KEY POINTS
- Black Forest Labs published FLUX.3 Image Klein to the Hugging Face Hub and ships open checkpoints that run in Hugging Face Diffusers.
- The model supports text-to-image and multi-reference image-to-image generation, adds bounding-box layout control for region placement and image edits, includes a distilled checkpoint (1024×1024 ~2s on an H100, 16 steps by default), quantized single-file transformer weights (bf16, fp8r, nvfp4), and LoRA/DreamBooth training scripts.
- Native Diffusers support, open checkpoints, and bounding-box layout control make FLUX.3 Image Klein immediately usable for research, fine‑tuning, and precise image composition workflows.
DECISION BRIEF
CONFIDENCE
WHAT CHANGED
Black Forest Labs published FLUX.3 Image Klein as open checkpoints that run natively in Hugging Face Diffusers. The release adds bounding-box layout control for region placement and image edits, provides a distilled checkpoint (1024×1024 images in ~2s on an H100, 16 steps by default), single-file transformer weights in bf16 and quantized formats (fp8r, nvfp4), LoRA/DreamBooth training scripts, and a modular Diffusers pipeline.
WHY NOW
Native Diffusers support plus open checkpoints and compact/quantized transformer files make the model immediately usable for generation, multi‑reference image‑to‑image edits, and fine‑tuning workflows. The distilled checkpoint’s fast generation on high‑end hardware and quantized/single‑file weights for smaller GPUs lower the barrier to local experimentation and integration into existing Diffusers‑based toolchains.
WHO IS AFFECTED
Users of Hugging Face Diffusers and Spaces, developers and researchers doing text‑to‑image and image‑to‑image work, people who fine‑tune models (LoRA/DreamBooth users), and practitioners seeking low‑VRAM local runs (quantized/INT8 pathways and single‑file weights). The model is also offered via BFL’s API, Playground, and integrations (ComfyUI), affecting those deployment paths.
CONFIRMED
• FLUX.3 Image Klein runs in Diffusers and is published on the Hugging Face Hub. • The model supports text‑to‑image and multi‑reference image‑to‑image generation. • The model adds bounding‑box layout control: prompts can include JSON regions with ids, bboxes ([y_min,x_min,y_max,x_max] on a 0–1000 scale), and descriptions; regions specify src_bbox and tgt_bbox for edits. • A distilled checkpoint generates 1024×1024 images in ~2 seconds on an H100 and uses 16 steps by default. • Transformer weights are provided as a single file in bf16 and two quantized formats (fp8r and nvfp4) and can be loaded with from_single_file; other components load by name. • LoRA training scripts for text‑to‑image and image editing are provided. • The Decoder reports FLUX.3 Image Klein is a 5.6‑billion‑parameter open‑weights model supporting up to ten reference images, native 4K output, pixel‑accurate editing, and bounding boxes for element placement. • The Decoder reports an INT8 build can run on NVIDIA GPUs with 8 GB VRAM, producing an image in under 9 seconds on an RTX 4070. • The Decoder reports the weights are released under the FLUX Kommunity License, with commercial use permitted for companies under $5 million annual revenue; the model is available via BFL API, Playground, Hugging Face Spaces, ComfyUI, and Diffusers.
UNCERTAIN
• Independent, third‑party benchmark verification of the model’s image quality, Elo claims, and speed across a range of hardware is not provided in these excerpts; The Decoder attributes Elo gains to Black Forest Labs’ internal testing. • Full terms and limitations of the FLUX Kommunity License beyond the summary in the Decoder excerpt are not included here; exact license text and edge cases are missing. • Resource requirements for training/fine‑tuning (beyond that the base checkpoint runs 50 steps by default and is intended for fine‑tuning) and real‑world latency on other GPU models/configurations are not fully specified in the excerpts. • The Hugging Face excerpt documents the API for bounding boxes and examples, but broader usability limits (failure modes, robustness of pixel‑accurate edits, and limits on number/size of references) are not detailed.
WHAT TO WATCH
Observable signals to monitor: (1) Hugging Face repo activity (downloads, commits, issues, Spaces usage) and sample galleries; (2) independent benchmark and Elo reports from third parties comparing FLUX.3 to other models; (3) community LoRA/DreamBooth checkpoints and training recipes published for this model; (4) integrations/posts showing INT8 or nvfp4 runs on consumer GPUs (timings on RTX 40-series and other cards); (5) clarifications or full text of the FLUX Kommunity License and any downstream enforcement or changes; (6) adoption/integration posts for ComfyUI, Diffusers pipelines, and toolkits (Ostris, NeMo) referenced in the excerpts.
WHY IT MATTERS
Native Diffusers support, open checkpoints, and bounding-box layout control make FLUX.3 Image Klein immediately usable for research, fine‑tuning, and precise image composition workflows.
EVIDENCE MAP
4Editorial claims linked to specific sources, with support, contradiction and context shown separately.
Black Forest Labs published FLUX.3 Image Klein on the Hugging Face Hub and it runs in Hugging Face Diffusers.
SUPPORTEDChecked 2026-10-15 · 2 supporting
The Hugging Face blog post states FLUX.3 Image Klein 'runs in Diffusers from day one' and points to the model/collection on Hugging Face.
SUPPORTS The DecoderMEDIA · 2026-10-15The Decoder article lists Diffusers among the local/hosted ways the model is available.
The model supports text-to-image and multi-reference image-to-image generation.
SUPPORTEDChecked 2026-10-15 · 2 supporting
Hugging Face describes 'One model does text-to-image and multi-reference image-to-image.'
SUPPORTS The DecoderMEDIA · 2026-10-15The Decoder describes the model as for text-to-image generation and image editing, and mentions generation with multiple reference images.
FLUX.3 Image Klein adds bounding-box layout control for region placement and image edits (boxes are provided in prompts/JSON regions).
SUPPORTEDChecked 2026-10-15 · 2 supporting
Hugging Face explains adding 'layout control' where users write bounding boxes into prompts and JSON region lists to place or edit parts of the image.
SUPPORTS The DecoderMEDIA · 2026-10-15The Decoder reports bounding boxes are used for element placement and pixel‑accurate editing.
A distilled checkpoint produces 1024×1024 images (the Hugging Face post reports ~2 seconds on an H100) and defaults to 16 inference steps.
SUPPORTEDChecked 2026-10-15 · 1 supporting
SOURCES & TIMELINE
2Black Forest Labs FLUX.3 Image Klein is here, and it runs in Diffusers from day one. One model does text-to-image and multi-reference image-to-image, and it adds something new, the layout control . You write bounding boxes into the prompt to say where each part of the image should go, or which parts of a reference image to move. Both open checkpoints are in the FLUX.3 Image Klein collection : You can head on to the…
Black Forest Labs has released FLUX 3 Image Klein, a 5.6-billion-parameter open-weights model for text-to-image generation, image editing, and generation with up to ten reference images. It supports native 4K output, pixel-accurate editing, and bounding boxes for element placement. Black Forest Labs says it scores the highest Elo ratings among all open models in internal testing, beating closed competitors Ideogram V…