# BluePixel-Turbo, for ComfyUI

The files WAS Node Suite edits pictures with BluePixel-Turbo through, laid out as ComfyUI's own
`models` folder: `diffusion_models/` for BluePixel Model Loader, `vae/` for BluePixel VAE Loader.

## Attribution Notice

This FLUX Model is licensed by Black Forest Labs Inc. under the FLUX Non-Commercial License. Copyright Black Forest Labs Inc. IN NO EVENT SHALL BLACK FOREST LABS INC. BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.

BluePixel-Turbo is a modified version of FLUX.2 [klein] 9B, made by vivo.

The transformers in `diffusion_models/` are WAS Node Suite's modified versions of vivo's
BluePixel-Turbo: converted from vivo's float32 Diffusers release to Black Forest Labs' FLUX.2
layout at bfloat16, and quantised from that bfloat16 file, as described below.

The FLUX Non-Commercial License is in `diffusion_models/LICENSE-FLUX-Non-Commercial.md`, the same
text vivo publishes beside BluePixel-Turbo. Any right to use the transformers is granted directly
by Black Forest Labs Inc. to each recipient under that licence, for non-commercial purposes only.
Commercial use needs a licence from Black Forest Labs at <https://bfl.ai/licensing>. These files
are not an official product of Black Forest Labs Inc. or of vivo, and neither has endorsed,
approved or validated them.

## Files

| File | Upstream | Licence |
|---|---|---|
| `diffusion_models/bluepixel_turbo_bf16.safetensors` | [openvivo/BluePixel-Turbo](https://huggingface.co/openvivo/BluePixel-Turbo) at `872eb46babdfb56fd3c8380ff1ef15ddd2961979`, `transformer/`, converted | FLUX Non-Commercial License: `diffusion_models/LICENSE-FLUX-Non-Commercial.md` |
| `diffusion_models/bluepixel_turbo_int8_convrot.safetensors` | int8 ConvRot of `bluepixel_turbo_bf16.safetensors` | FLUX Non-Commercial License |
| `diffusion_models/bluepixel_turbo_w6a8.safetensors` | W6A8 of `bluepixel_turbo_bf16.safetensors` | FLUX Non-Commercial License |
| `diffusion_models/bluepixel_turbo_w4a8.safetensors` | W4A8 of `bluepixel_turbo_bf16.safetensors` | FLUX Non-Commercial License |
| `vae/dc-ae-f64c128-mix-1.0.safetensors` | [mit-han-lab/dc-ae-f64c128-mix-1.0-diffusers](https://huggingface.co/mit-han-lab/dc-ae-f64c128-mix-1.0-diffusers) at `a735dcafe414e537f3162e875144ae8405bdebdf`, `diffusion_pytorch_model.safetensors`, byte for byte; the same file as vivo's `vae/` | Apache-2.0, MIT Han Lab: `vae/LICENSE-DC-AE` |

## How the transformers were made

`bluepixel_turbo_bf16.safetensors` holds vivo's 233 Diffusers tensors as Black Forest Labs' 201
FLUX.2 tensors: each block's query, key and value joined into one `qkv`, the single blocks'
projections renamed, the final modulation's two halves put in shift-then-scale order, and every
value rounded to the nearest bfloat16. ComfyUI's FLUX.2 model reads it as it reads FLUX.2 Klein 9B.

| Suffix | Format |
|---|---|
| `_bf16` | bfloat16 |
| `_int8_convrot` | `int8_tensorwise` with ConvRot rotation, group 256, one scale per output row, chosen for the lowest error |
| `_w6a8` | `w6a8_int8`, 6-bit weights in groups of 32 with ConvRot rotation, group 256, run with int8 activations |
| `_w4a8` | `asym_w4a8_int8`, 4-bit weights in groups of 16 with a codebook and ConvRot rotation, group 256, run with int8 activations |

Quantised: the 112 linear layers of the 8 double-stream and 24 single-stream blocks. The input,
text, time, modulation and final layers and every norm are bfloat16 in every file.

vivo's transformer shards, SHA-256:

    194f13c1678f274138cf6b82bc621a2b0f38c6733b67f853540460e12d5debc5  transformer/diffusion_pytorch_model-00001-of-00010.safetensors
    18a070fe1f7b13b8ecea7b9d565d510845f88c3122d8ecc008dd9023019c6410  transformer/diffusion_pytorch_model-00002-of-00010.safetensors
    e1d2371c6ea16c5d08acf04cab26281178c59b529913b8883f76ba3401416191  transformer/diffusion_pytorch_model-00003-of-00010.safetensors
    c50613fcb3115ce3466649d9e1dd66329491d70bc2f0ea1267bc30c4dcd9e174  transformer/diffusion_pytorch_model-00004-of-00010.safetensors
    c2f32df7e0560384464df7db681fdcea81e101d4fe18622aa0f488ff3d870e21  transformer/diffusion_pytorch_model-00005-of-00010.safetensors
    269a7aff70c4ff257560a781a93dad6795ddba446a1b721b1091bf95fa6e5bd7  transformer/diffusion_pytorch_model-00006-of-00010.safetensors
    72b284ea2debf68d587420f95396124a85a46da5594cadedb3d78c3c31368c10  transformer/diffusion_pytorch_model-00007-of-00010.safetensors
    5325644990aa6a3060a220dd8cee4cad381222be0a64bc5c44e048d320aa76e0  transformer/diffusion_pytorch_model-00008-of-00010.safetensors
    3695b09b2a1d2a0682c50413833839f7e827d759f47ea218d1a5a13c8c583c90  transformer/diffusion_pytorch_model-00009-of-00010.safetensors
    8107a6f0aa435a3c5b3245c6c8ee843a74a9331e85a8336039e7319c08eeccf4  transformer/diffusion_pytorch_model-00010-of-00010.safetensors

## SHA-256

    f4bf27e01173ad9b6fc2580f91ec1ac01bda3e32147f3252bd2867d4520867d3  diffusion_models/bluepixel_turbo_bf16.safetensors
    ad2b059d66a48b55bfd32adb2301c677fc4684837c76b68f751f53c0e0a8f88e  diffusion_models/bluepixel_turbo_int8_convrot.safetensors
    2f7bf1e21c099b5c852c16442319ab7f7035b86ff6e06a82d119229340f3493d  diffusion_models/bluepixel_turbo_w6a8.safetensors
    118efb37b1792bb7fe0500081d531f0d436ee9d29e32b566767493dce9b35c97  diffusion_models/bluepixel_turbo_w4a8.safetensors
    ae11d521da6370b9dec34758f78a6359a3d486b121afb460bd549e8182c518d1  vae/dc-ae-f64c128-mix-1.0.safetensors

## Not mirrored here

| File | Where it is |
|---|---|
| `qwen_3_8b.safetensors`, the text encoder | [Comfy-Org/Flux-2-Klein-9B](https://huggingface.co/Comfy-Org/Flux-2-Klein-9B), `split_files/text_encoders/`. Its 399 tensors are byte for byte the ones in vivo's `text_encoder/` |