Comfy-Org Makes Black Forest Labs' FLUX.2 Klein Run in 1.2 Seconds Locally
Comfy-Org repackaged the FLUX.2 klein 4B text encoder and VAE for drop-in ComfyUI use, pushing the sub-second open model past 800k downloads.
- Comfy-Org repackaged FLUX.2 klein 4B as single-file safetensors for ComfyUI, surpassing 800k downloads.
- Bundle includes the 336 MB FLUX.2 VAE, Qwen3-4B text encoder (bf16 and fp4), and both distilled and base transformers.
- Underlying FLUX.2 klein is a 4B rectified flow transformer with sub-second inference on consumer GPUs.
- Distilled 4-step model runs in ~1.2s on an RTX 5090 using 8.4 GB VRAM; base variant takes ~17s.
- Apache 2.0 license on the 4B weights allows royalty-free commercial use.
- Unified single checkpoint handles text-to-image, image editing, and multi-reference composition.
FLUX.2 klein 4B gets a cleaner ComfyUI install
A Comfy-Org package has made Black Forest Labs’ smallest FLUX.2 image model easier to run locally. It converts the VAE, text encoder, and diffusion transformer into standalone safetensors files that fit ComfyUI’s standard model directories. At publication, Hugging Face showed more than 800,000 downloads and 3,659 likes.
The underlying weights come from the upstream 4B release. Black Forest Labs distributes that release as a Diffusers-style directory containing weights, configuration files, and other pipeline metadata. Comfy-Org reorganizes those components for ComfyUI’s native loaders, removing the need to convert or remap the files manually.
One package, three moving parts
The complete repository occupies 27.7 GB because it includes two diffusion checkpoints and two versions of the text encoder. A working installation needs one diffusion checkpoint, one text encoder, and the VAE.
| File | Purpose |
|---|---|
flux2-vae.safetensors |
The 336 MB variational autoencoder that converts between image pixels and the model’s compressed latent representation. |
qwen_3_4b.safetensors |
The 8.04 GB Qwen3-4B prompt encoder stored in BF16 precision. |
qwen_3_4b_fp4_flux2.safetensors |
An FP4-quantized prompt encoder that reduces storage and memory pressure. |
flux-2-klein-4b.safetensors |
The 7.75 GB distilled transformer, tuned for four-step generation. |
flux-2-klein-base-4b.safetensors |
The 7.75 GB base transformer, intended for longer sampling schedules, fine-tuning, and custom control pipelines. |
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
