FlowDIS Int8 Convrot is a quantized version of the FlowDIS transformer model designed for image segmentation tasks. 03673. The model is distributed in a quantized checkpoint format compatible with ComfyUI and was produced using the convert_to_quant process.
This pre-quantized model is tailored for use with Any Prompt DIS, particularly supporting low-VRAM operation modes such as --int8 and --t5-int4. 4 GiB at 512²), making it suitable for users with 24 GB graphics cards. In contrast, the bf16 pipeline requires significantly more memory (35 GiB peak), which exceeds the capacity of such cards. The model specifically addresses the challenge that quantizing the checkpoint locally would require loading the bf16 transformer, a process that is not feasible on a 24 GB card, thus necessitating this pre-quantized upload.
The technical details include the quantization of 228 double_blocks and single_blocks linear layers to INT8, using row-wise scales, a ConvRot group size of 256, and pre-applied rotations to weights. The model is provided under the picsart-flowdis-model-license. It is available for download and use through Hugging Face.
evidence_sufficient": true}
In the Other AI space, FlowDIS Int8 Convrot takes a focused approach. It focuses on providing a quantized, open-source transformer model for efficient image segmentation tasks. FlowDIS Int8 Convrot is an open-source project aimed at computer vision researchers. The project is open source (Open Source). It runs on the command line.
Behind FlowDIS Int8 Convrot is Albertchen96, and the product first shipped in 2026. The project is developed in the open on GitHub with 13 commits in the last 90 days. Among its 4 catalogued features are image segmentation, INT8 quantization, and transformer architecture.
Latest indexed changes and source events
Albertchen96/FlowDIS-int8-convrot verified by the PulseGate indexer
Other apps tracked under the same category.
