BLOCKv0.6 · ComfyUI format
This is AliceKJ/BLOCKv0.6 repackaged so it loads in ComfyUI.
The weights are unchanged; only the tensor names were converted from the diffusers layout to the
native FLUX.2 layout that ComfyUI's UNETLoader reads (fused QKV, double_blocks / single_blocks,
BFL order for the final adaLN shift/scale). ComfyUI can't convert diffusers-format FLUX.2 weights on
its own, so the original file doesn't load there.
block-v0.6-flux2-klein-4b.safetensors· transformer only, bf16, 149 tensors. Names and shapes matchflux-2-klein-base-4b.safetensorsfrom Comfy-Org/flux2-klein.
Using it in ComfyUI
Put the file in models/diffusion_models and use it where the FLUX.2 Klein 4B base image-edit
template loads its UNET. Pair it with:
- text encoder:
qwen_3_4b.safetensors(CLIPLoader, typeflux2) - VAE:
flux2-vae.safetensors
Settings from the original model card: the input is a 512x512 front/back Minecraft preview, 30 steps, CFG 4.0, and this prompt:
Image-to-image translation using the reference image. The reference shows the same 3D Minecraft character with front and back views in a single image. Generate the corresponding Minecraft skin UV atlas in 64x64 pixel-art UV layout. High-quality anime-style. Flat shading, sharp pixel edges, no blur, no anti-aliasing. Keep consistent UV placement and mapping; match the same character design from the reference. Model type: classic (auto-detected Minecraft player model).
The output is a 512x512 atlas on a white background. Downsample it 8x (nearest) and rebuild the alpha from the Minecraft UV layout to get a usable 64x64 skin.
Credits and license
All credit for the model goes to the BLOCK authors. Apache 2.0, same as the original.
@article{guo2026block,
title={BLOCK: An Open-Source Bi-Stage MLLM Character-to-Skin Pipeline for Minecraft},
author={Guo, Hengquan},
journal={arXiv preprint arXiv:2603.03964},
year={2026},
url={http://arxiv.org/abs/2603.03964}
}