ImageKleinStart creating1. Get access to the weights
- Sign in to Hugging Face and open black-forest-labs/flux-3-image-klein.
- Accept the FLUX Kommunity License on the page (the repo is gated).
- Pick the checkpoint: flux-3-image-klein (distilled, 16 steps — for making images) or flux-3-image-klein-base (50 steps — for fine-tuning and LoRA training). ComfyUI users can use flux-3-image-klein-comfy.
2a. ComfyUI
ComfyUI supports FLUX 3 Image Klein from day one. Update ComfyUI, download the model files from the flux-3-image-klein-comfy repo, and load the FLUX 3 Image Klein template. On an 8 GB card, use the INT8 build — that is the configuration BFL quotes for 8 GB GeForce RTX GPUs.
2b. Hugging Face Diffusers
Install Diffusers from source (support landed with the release), then:
pip install git+https://github.com/huggingface/diffusers.git
import torch
from diffusers import ModularPipeline
pipe = ModularPipeline.from_pretrained("black-forest-labs/flux-3-image-klein")
pipe.load_components(dtype=torch.bfloat16)
pipe.to("cuda")
# text-to-image
image = pipe(
prompt="A wide-angle photo of a red vintage tram on a rainy Lisbon street at dusk, "
"warm shop lights reflected on wet cobblestones, soft mist, muted palette.",
generator=torch.Generator().manual_seed(0),
output="images",
)[0]
image.save("tram.png")
# edit the same image
night = pipe(
prompt="The same tram on the same street, now at night under heavy snow.",
image=image,
generator=torch.Generator().manual_seed(0),
output="images",
)[0]
night.save("tram_snow.png")The pattern (ModularPipeline, image= for edits) follows the official model card; the prompts here are our own.
2c. Reference code
Black Forest Labs' GitHub repo black-forest-labs/flux-3-image-klein has a CLI (scripts/cli.py --model distilled, weights under weights/) and a Gradio demo for drawing bounding boxes.
How much VRAM do you need?
| Setup | GPU memory |
|---|---|
| INT8 build in ComfyUI or Diffusers | as little as 8 GB (GeForce RTX) |
| Reference code, NVFP4, 1024×1024 | ≈25 GB peak |
| Reference code, FP8, 1024×1024 | ≈26 GB peak |
| Reference code, BF16, 1024×1024 | ≈31 GB peak |
| + prompt upsampler | ≈28 GB more |
The reference-code figures are from the model card (measured on a B200). The 8 GB figure is BFL's for the INT8 build on GeForce RTX; BFL hasn't published speed for an 8 GB card or for 4K output.
How fast is it?
| GPU | Build | 1 MP text-to-image |
|---|---|---|
| RTX 4070 | distilled, INT8 | 8.9 s (edit: 18.9 s) |
| RTX 5090 | distilled, INT8 | 4.1 s (edit: 7.2 s) |
| RTX 5090 | base (50 steps), INT8 | 12.5 s (edit: 27.4 s) |
| DGX Spark | base (50 steps) | 43.8 s |
| GB300 / DGX Station | base (50 steps) | 1.7 s |
| GB300 | distilled, NVFP4 | 0.55 s |
GPU time only (no prompt upsampling), from NVIDIA measurements published by BFL and VentureBeat's reading of the charts. At 2048×2048 an RTX 5090 takes about 15 s and an RTX 4070 about 50 s with the distilled INT8 build.
Fine-tuning a LoRA
Train on flux-3-image-klein-base. Ostris's AI Toolkit and Diffusers support LoRA training from day one, and NVIDIA NeMo AutoModel has recipes that work on the released checkpoints. Serving a LoRA costs about 5% extra generation time.