Welcome FLUX.3 Image Klein in Diffusers
- ID
- 35878
- Status
- summarized
- Published
- 15 Oct 2026, 8:00 AM
- Fetched
- 16 Oct 2026, 12:34 AM
- Provider
- Hugging Face Blog
- Category
- developer-ai
- Original URL
- https://huggingface.co/blog/flux-3
- Source URL
- https://huggingface.co/blog/feed.xml
Summary
- Score
- 7.0
- Created
- 16 Oct 2026, 12:34 AM
- Tags
- Audience
- developersai_ml_learnersvibe_coders
What happened
Black Forest Labs released FLUX.3 Image Klein, and Hugging Face shipped Diffusers support on day one. One model handles text-to-image, multi-reference image-to-image, and a new layout control where you write bounding boxes into the prompt to place image parts or move regions of a reference image. Two open checkpoints are published: black-forest-labs/flux-3-image-klein (distilled, 16 steps, no CFG) and flux-3-image-klein-base (50 steps, CFG), with the distilled version generating a 1024x1024 image in about 2 seconds on an H100. NVFP4 and FP8 weights load via from_single_file for smaller GPUs, LoRA training scripts exist for text-to-image and image editing, and it ships as a Modular Diffusers pipeline. It needs the latest diffusers and transformers (for the Qwen3-VL text encoder).
Why it matters
If you self-host image generation, the practical decision here is checkpoint choice and VRAM: the distilled checkpoint runs 16 steps with no CFG at roughly 2 seconds per 1024x1024 on an H100, and FP8/NVFP4 weights load through from_single_file, so you can test on a much smaller GPU before committing to H100-class rental. The bounding-box layout control also means you may not need a separate ControlNet-style conditioning stack for simple placement or region-move edits. Test the distilled checkpoint on your actual target GPU first; the base model's 50 steps with CFG is a different cost profile entirely.
Discussion angle
Prompt-level bounding boxes for layout control versus a separate ControlNet-style pipeline - is writing coordinates into the prompt actually usable in production, or does it break down on multi-object scenes? Worth testing live against the same prompt with and without boxes.