Comfy-Org Packages FLUX.2 Klein 9B Into One 18GB Download
Black Forest Labs' distilled 9B image model lands in ComfyUI as a single safetensors file, generating and editing images in under a second on consumer GPUs.
Takeaways
- Comfy-Org repackaged FLUX.2 klein 9B as a single safetensors for ComfyUI, 800K+ downloads.
- Distilled 9B runs in 4 steps, sub-second on GB200, ~2s on RTX 5090, ~19.6 GB VRAM.
- Unifies text-to-image, single-reference editing, and multi-reference composition in one model.
- GGUF Q8_0 quants from Unsloth bring it down to ~10 GB for 8-12 GB cards.
- License is FLUX Non-Commercial; the 4B sibling is the Apache 2.0 option for commercial use.
- Weak at legible text and factual accuracy, prompt adherence sensitive to phrasing.
Comfy-Org has published Black Forest Labs’ FLUX.2 klein 9B weights as an 18.2 GB .safetensors diffusion checkpoint for ComfyUI. The release reduces model installation to one download and provides an official workflow for generation and editing. Architecture, runtime requirements, and licensing remain tied to BFL’s original 9B repository.
The ComfyUI checkpoint has passed 800,000 downloads on Hugging Face at publication time. That counter records file requests, including repeat downloads, rather than unique installations.
One file cuts setup
Comfy-Org’s package contains the distilled 9-billion-parameter diffusion model in a format supported by ComfyUI’s model loader. A current ComfyUI build and the official template are still required to supply the correct nodes, conditioning, and decoding pipeline.
- Model: FLUX.2 klein 9B distilled
- File:
flux-2-klein-9b.safetensors - Size: approximately 18.2 GB
- Tasks: text-to-image generation, image editing, and multi-reference composition
- License: FLUX Non-Commercial License
The underlying model is a rectified-flow transformer. Rectified flow trains a model to move samples from noise toward an image along relatively direct trajectories. Its four-step inference comes primarily from step distillation, which teaches the checkpoint to approximate a longer sampling process with fewer transformer passes.
Four passes drive the speed
Both distilled klein sizes target four inference steps. Black Forest Labs reports about 0.5 seconds per image on an NVIDIA GB200 and roughly 2 seconds on an RTX 5090 for the distilled 9B model, with approximately 19.6 GB of VRAM use. The company reports about 35 seconds on an RTX 5090 for the undistilled 9B base model.
| Variant | Sampling | Published performance | Primary use |
|---|---|---|---|
| Distilled 9B | 4 steps | About 0.5 seconds on GB200 and 2 seconds on RTX 5090 | Low-latency inference |
| Base 9B | Undistilled schedule |