[GH-ISSUE #273] Integrating EntroPack for flexible-bitrate weight compression #71

New issue

Open

opened 2026-10-26 22:19:57 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 22:19:57 +00:00

Owner

Originally created by @mi804 on GitHub (Oct 23, 2026).
Original GitHub issue: https://github.com/Comfy-Org/comfy-kitchen/issues/273

Hi @comfyanonymous,

I'm one of the authors of EntroPack, an open-source PyTorch tensor compression library. Would you be interested in supporting EntroPack in comfy-kitchen?

EntroPack makes it easy to quantize model weights to a chosen storage bitrate, including fractional targets. For ComfyUI, this could mean offering quantized models at any target between 3 and 8 bits per parameter, giving users more choices to fit their VRAM budget while preserving model quality. It also works on top of existing INT8 quantization: INT8 weights can be compressed further for storage and restored to INT8 for inference.

We've tested EntroPack on a range of diffusion models, including MOSS-Image, and released pre-quantized weights for the community to use.

We'd like to bring this option to comfy-kitchen and would be happy to contribute a PR. We're open to either integrating the relevant low-level code directly or using EntroPack as a dependency, whichever fits the project better. If you're interested, please let us know which approach you'd prefer.

Thanks for your work on comfy-kitchen!

Paper: https://arxiv.org/abs/2609.34185

GitHub: https://github.com/modelscope/entropack

Originally created by @mi804 on GitHub (Oct 23, 2026). Original GitHub issue: https://github.com/Comfy-Org/comfy-kitchen/issues/273 Hi @comfyanonymous, I'm one of the authors of [EntroPack](https://github.com/modelscope/entropack), an open-source PyTorch tensor compression library. Would you be interested in supporting EntroPack in comfy-kitchen? EntroPack makes it easy to quantize model weights to a chosen storage bitrate, including fractional targets. For ComfyUI, this could mean offering quantized models at any target between 3 and 8 bits per parameter, giving users more choices to fit their VRAM budget while preserving model quality. It also works on top of existing INT8 quantization: INT8 weights can be compressed further for storage and restored to INT8 for inference. We've tested EntroPack on a range of diffusion models, including MOSS-Image, and released [pre-quantized weights](https://www.modelscope.cn/collections/DiffSynth-Studio/EntroPack-Quantized-Models) for the community to use. We'd like to bring this option to comfy-kitchen and would be happy to contribute a PR. We're open to either integrating the relevant low-level code directly or using EntroPack as a dependency, whichever fits the project better. If you're interested, please let us know which approach you'd prefer. Thanks for your work on comfy-kitchen! Paper: https://arxiv.org/abs/2609.34185 GitHub: https://github.com/modelscope/entropack

No labels

pull-request

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

Comfy-Org/comfy-kitchen#71

No description provided.

Rows
Columns