socialtrait/flux-3-image-klein-infendpoint

Socialtrait FLUX.3 Klein deployment checkpoint

BF16 deployment subset of black-forest-labs/flux-3-image-klein, upstream commit 337337621b8526e120f2275d250242f4a6bedd75. Weights are unchanged. Includes Diffusers transformer, text encoder, VAE, configuration, tokenizers, original licenses and notices. Duplicate original weights, FP8/NVFP4 alternatives and gallery assets are omitted. See README.upstream.md for the original upstream model card.

handler.py provides the HF Inference Endpoints custom handler. modal_app.py deploys an authenticated API on Modal; see DEPLOYMENT.txt. Defaults: 16 steps, 1024 x 1024, BF16 transformer and text encoder, FP32 VAE, no CPU offloading, no prompt upsampling. The separate prompt-upsampler model is not included. A10 GPU memory fit and throughput still require benchmarking. The optional GGUF fallback requires a separate Unsloth checkpoint.

The /sdxl Modal route accepts the existing persona generator payload and returns base64 JPEG. SDXL-only options are ignored; this route uses 16 steps. Use /generate for explicit dimensions and 4–16 steps.

This is a deployment subset, not a new model or a change to the upstream license.

Rights to use these weights are granted directly by Black Forest Labs under LICENSE.md. See the required attribution in NOTICE.md. This deployment packaging is maintained by Socialtrait and is not an official Black Forest Labs product. Model weights are unchanged.