FlowDIS-int8-convrot

FlowDIS transformer — INT8 ConvRot

INT8 quantization of the FlowDIS DiT (12B, FLUX.1-schnell architecture) with group-wise Hadamard rotations (ConvRot, arXiv:2512.03673), in the ComfyUI-native quantized checkpoint format produced by convert_to_quant.

Made for Any Prompt DIS, whose low-VRAM mode (--int8 --t5-int4) runs the full segmentation pipeline in ~21 GiB peak VRAM at 1024² (~18.4 GiB at 512²) — it fits a 24 GB card, which the bf16 pipeline (~35 GiB peak) does not. Quantizing this checkpoint yourself requires loading the bf16 transformer, so a 24 GB card also cannot produce it locally — hence this pre-quantized upload.

Details

• 228 double_blocks/single_blocks linear layers quantized to INT8 (row-wise scales, ConvRot group size 256, rotations pre-applied to weights); input/modulation/final layers kept in bf16.
• End-to-end relative error vs bf16: ~1.2% on mask latents; visually indistinguishable masks in our tests.
• Runtime: flowdis/quant.py in the repo above — uses the fused comfy-kitchen INT8 kernel when available (quantized linears ~1.5× faster than bf16), otherwise a torch.compile fallback.

Reproduce from the bf16 checkpoint:

pip install convert-to-quant
ctq -i flowdis-transformer.safetensors \
    -o flowdis-transformer-int8-convrot.safetensors \
    --comfy_quant --int8 --convrot --convrot-group-size 256 \
    --exclude-layers "img_in|txt_in|time_in|vector_in|mod|final_layer" \
    --save-quant-metadata

Usage

# place it next to the other FlowDIS weights:
hf download Albertchen96/FlowDIS-int8-convrot flowdis-transformer-int8-convrot.safetensors \
    --local-dir <root_model_dir>

# then, in the any-prompt-dis repo:
python inference_si.py --root-model-dir <root_model_dir> --int8 \
    --image-path input.jpg --prompt "dog" --output-path mask.png

License

Derivative of the FlowDIS weights by Picsart AI Research; distributed under the same PicsArt Inc. FlowDIS Model License (non-commercial). Review it before redistribution or commercial use.