Welcome FLUX.3 Image Klein in Diffusers
Black Forest Labs released FLUX.3 Image Klein, a Diffusers‑compatible model that supports text‑to‑image, multi‑reference image‑to‑image, and bounding‑box layout control. The distilled checkpoint generates a 1024 × 1024 image in about 2 seconds on an H100 (16 inference steps by default), while the base checkpoint runs 50 steps with classifier‑free guidance and is the target for LoRA fine‑tuning. Quantized variants (fp8r and nvfp4) can be loaded with `from_single_file`, and the repo provides LoRA training scripts for both generation and editing tasks.
⚡ Key Takeaways
- Distilled checkpoint: 1024 × 1024 image in ~2 s on H100 (16 steps default).
- Supports layout control via JSON bounding boxes on a 0‑1000 normalized scale.
- Quantized formats fp8r (row‑wise FP8) and nvfp4 load with `from_single_file`.
- LoRA training scripts are included for text‑to‑image and image‑to‑image fine‑tuning.
- Memory‑saving flags: `--gradient_checkpointing`, `--cache_latents`, `--use_8bit_adam`.
🔧 Tools & Libraries
Engineers can drop‑in FLUX.3 Image Klein into existing Diffusers pipelines to get fast, high‑resolution generation with precise spatial control, and immediately fine‑tune the model using provided LoRA scripts without needing a custom training stack.
✅ Practical Steps
- Install the latest `diffusers` and `transformers` (for the Qwen3‑VL text encoder) and load the distilled checkpoint via `from_single_file`.
- Generate an image by passing a prompt, optional reference images, and a JSON list of layout regions; adjust `height`, `width`, or `num_inference_steps` as needed.
- Fine‑tune the base checkpoint with the provided LoRA training scripts, optionally adding `--gradient_checkpointing`, `--cache_latents`, and `--use_8bit_adam` to reduce VRAM usage.
Want the full story? Read the original article.
Read on Hugging Face Blog ↗