You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This is a modified copy of black-forest-labs/flux-3-image-klein and is distributed under the FLUX Kommunity License (LICENSE.md in this repository): non-commercial use, with a commercial-output exception for Qualifying Users. Your rights to use it are granted directly by Black Forest Labs Inc. under that license.
Log in or Sign Up to review the conditions and access this model content.
FLUX.3 [klein], FP8 — radiance container
black-forest-labs/flux-3-image-klein — text to image, editing and generation from up to ten reference
pictures — as a single .rad container for the
radiance inference engine (AMD RDNA4, ROCm).
This is not an official Black Forest Labs release and is not endorsed by Black Forest Labs.
| File | flux-3-klein-fp8.rad — 9.32 GiB |
| Weights | the text encoder's (Qwen3-VL-4B-Instruct) and the diffusion transformer's linears FP8 E4M3 per 128×128 block (bf16 scales); the small linears, the norms and the VAE bf16 |
| Steps | 16, unguided (the distilled checkpoint) |
| Needs | radiance built from main after 1.8.1 — FLUX.3 [klein] support is not in 1.8.1 and comes with the next release |
Serve
radiance --model flux-3-klein-fp8.rad --host 0.0.0.0 --port 8000
The server speaks OpenAI's Images API: POST /v1/images/generations and POST /v1/images/edits
(JSON, or multipart with image / image[], as OpenAI's SDKs send them; each reference is shrunk to
1024x1024 of area when larger). Pictures come back as b64_json (png, jpeg or webp). Sizes up to 2048x2048,
sides multiples of 16, 1024x1024 by default; seed, n, streaming with partial_images. RGB only
(background: transparent is refused).
4.4 seconds a 1024x1024 picture on two Radeon AI PRO R9700 (gfx1201).
Against diffusers' own bf16 pipeline from the same noise (1024x1024, seed 7), PSNR 34.4 dB.
Prompts
The model takes full scene captions. Black Forest Labs recommend running every request through a prompt
upsampler first (flux-3-image-klein-pup, or
any language model with their system_prompts.json); radiance reads the prompt as given.
How this file was made
rad-convert black-forest-labs/flux-3-image-klein --recipe flux-3-klein-fp8.recipe -o flux-3-klein-fp8.rad
black-forest-labs/flux-3-image-klein is the diffusers download of that repository. The recipe (flux-3-klein-fp8.recipe in this
repository) names what is quantised; everything it does not name is the checkpoint's own dtype:
te.blk.*.attn_qkv.weight rtn codes=fp8_e4m3 block=128x128 scale=bf16
te.blk.*.attn_output.weight rtn codes=fp8_e4m3 block=128x128 scale=bf16
te.blk.*.ffn_gate_up.weight rtn codes=fp8_e4m3 block=128x128 scale=bf16
te.blk.*.ffn_down.weight rtn codes=fp8_e4m3 block=128x128 scale=bf16
dit.mode.*.attn_qkv.weight rtn codes=fp8_e4m3 block=128x128 scale=bf16
dit.mode.*.attn_output.weight rtn codes=fp8_e4m3 block=128x128 scale=bf16
dit.mode.*.ffn_in.weight rtn codes=fp8_e4m3 block=128x128 scale=bf16
dit.mode.*.ffn_out.weight rtn codes=fp8_e4m3 block=128x128 scale=bf16
dit.single.*.attn_qkv.weight rtn codes=fp8_e4m3 block=128x128 scale=bf16
dit.single.*.attn_output.weight rtn codes=fp8_e4m3 block=128x128 scale=bf16
dit.single.*.ffn_in.weight rtn codes=fp8_e4m3 block=128x128 scale=bf16
dit.single.*.ffn_out.weight rtn codes=fp8_e4m3 block=128x128 scale=bf16
vae.* cast dtype=bf16
License
This FLUX Model is licensed by Black Forest Labs Inc. under the FLUX Kommunity License (LICENSE.md), and
NOTICE carries its attribution notice. Any rights to use this model are granted directly by Black Forest
Labs Inc. to you under that license. This container is a modified copy: its linear weights are re-encoded as
FP8 (see above). The text encoder inside it is Qwen3-VL-4B-Instruct, under the Apache License 2.0
(LICENSE-Qwen3-VL).
Model tree for StillDeadcode/radiance-flux-3-klein-fp8
Base model
black-forest-labs/flux-3-image-klein