---
license: apache-2.0
base_model: black-forest-labs/FLUX.2-klein-4B
library_name: onnxruntime-web
pipeline_tag: text-to-image
tags:
  - flux
  - webgpu
  - onnx
  - text-to-image
  - browser
---

# FLUX.2 Klein 4B WebGPU

This is a community conversion of [black-forest-labs/FLUX.2-klein-4B](https://huggingface.co/black-forest-labs/FLUX.2-klein-4B)
at revision `e7b7dc27f91deacad38e78976d1f2b499d76a294`. It packages the complete 1024×1024, four-step
text-to-image path for direct browser inference with ONNX Runtime WebGPU.
It is not a Diffusers or Transformers checkpoint.

## Contents

The release contains the full Qwen text encoder, four FP16 transformer
partitions, the FP16 VAE decoder, and immutable JSON manifests. The active
payload is approximately 13.7 GiB across 378 files. Every binary referenced by
the manifests has a SHA-256 and byte-length record.

The entry manifest is:

```text
models/klein-4b/pipeline-1024/manifest.json
```

Manifest resource paths are relative to the manifest that declares them. A
consumer must resolve those paths against the manifest URL before fetching.
The companion browser app accepts an alternate entry manifest with its
`manifest` query parameter.

## Conversion and modification disclosure

The source BF16 model was exported to FP16 ONNX with external data. Weights are
full and unquantized. To avoid FP16 saturation in ONNX Runtime WebGPU during
the second denoising step, the 25 residual `Clip` bounds were changed from
±65504 to ±60000. This is an activation-graph modification, not a weight
quantization or a LoRA.

The altered transformer partitions passed native ONNX Runtime validation against
the saved source fixture with a minimum cosine similarity of 0.9999966. The
complete browser pipeline was rendered on Apple WebGPU with finite latents and
finite decoded pixels. On the validation machine, a warm 1024×1024 four-step
generation took 26.41 seconds. Detailed reports are in `validation/`.

## Requirements and limitations

- A browser with WebGPU and enough GPU/process memory for the complete model.
- This release was designed for a 16 GB target, but its process high-water mark
  has not yet been measured across target browsers.
- It exposes a fixed 1024×1024, four-step prompt path; safety classification,
  negative prompts, and general user-facing controls are not included.

## Licence and attribution

The upstream FLUX.2 Klein 4B source is Apache-2.0. This derivative includes the
upstream `LICENSE.md` and retains the upstream model identity and revision.
Black Forest Labs and FLUX are trademarks of their respective owners; this
community conversion is not affiliated with or endorsed by Black Forest Labs.