t8star/BluePixel-Comfy

BluePixel for ComfyUI

English | 简体中文

Three standalone components converted from openvivo/BluePixel-Turbo for native ComfyUI image editing. Use them with ComfyUI BluePixel T8, which provides the loaders and eight editable canvas workflows.

Download

Place each file in the matching folder under your ComfyUI installation. Total download: 28.19 GB (decimal).

Component	File	ComfyUI folder	Size
Diffusion · INT8 ConvRot	bluepixel_turbo_int8_convrot.safetensors	models/diffusion_models/	9.09 GB
Qwen3-8B · BF16	bluepixel_qwen3_8b_original.safetensors	models/text_encoders/	16.39 GB
DCAE-F64 · FP32	bluepixel_dcae_f64.safetensors	models/vae/	2.71 GB

Each component is one Safetensors file. The text encoder includes its tokenizer. File sizes and SHA256 hashes are recorded in checksums.json; adjacent manifests document conversion mappings and source hashes.

Quick start

1. Install BluePixel T8 through ComfyUI Manager, or follow the manual installation guide.
2. Download all three components above and restart ComfyUI.
3. Download the JSON for a 1K preview or 4K editing workflow, select the files and input image, and enter an editing instruction. Open the local canvas JSON with Ctrl+O, or drop it onto ComfyUI.
4. Start with 4 steps, CFG 3. The 4K example outputs 4096×4096.

Requires Python 3.10+, ComfyUI 0.43.0+, Diffusers 0.40+ and a compatible PyTorch / comfy-kitchen CUDA environment. Use the package's three BluePixel Loader nodes, then connect their standard outputs to ComfyUI sampling and VAE nodes. DCAE-F64 has 64× spatial compression and cannot be replaced by an ordinary Flux2 VAE.

Conversion and validation

• Source weights: openvivo/BluePixel-Turbo at e9c7a641a4bcd9851eef50c1ce9b07fa44f6044c.
• Diffusion weights are mapped to the native Flux2/Klein layout and quantized with INT8 ConvRot. Quantization is lossy; this is not a BF16-equivalent checkpoint.
• Qwen3 retains its original precision; DCAE is stored in FP32. No additional training or LoRA is applied to this release.
• Initial model validation included 107 unit tests and eight complete inference paths, an actual 4K edit and text-encoder / VAE comparisons. See validation conditions and measured differences.

Speed, memory requirements and output quality depend on hardware and task. Download size is not a VRAM requirement. These files are intended for the linked ComfyUI integration; they are not a drop-in Diffusers pipeline directory.

License and attribution

Weights retain their component-specific upstream terms. The diffusion model derives from FLUX.2 Klein 9B and is subject to the FLUX Non-Commercial License. The upstream BluePixel repository also supplies an OpenAtom model license; Qwen3 and DCAE are attributed under Apache-2.0. The upstream model card's Apache metadata does not replace the base-model restrictions.

Read LICENSE.md, NOTICE and the preserved files in licenses before use. This conversion does not grant additional commercial rights. The node repository's GPL license applies to its code, not these weights.

Credits: vivo BluePixel, Black Forest Labs, Qwen, NVIDIA DCAE, ComfyUI and comfy-kitchen. Converted and packaged by T8star.

T8star

Bilibili · YouTube · API · Free gallery · Online AI apps · ComfyUI bundle · Hugging Face