The Ling-mini-2.0 text encoders that the Ming-Image-0.1-Design models run on, hosted separately so any Ming Image checkpoint - ours or a community finetune - can point at them instead of shipping its own copy.
Originally released by inclusionAI on Hugging Face under the MIT license, and repackaged as single files for ComfyUI by Comfy-Org. All credit for the weights goes to the inclusionAI team.
Two versions, and they are not interchangeable
The base checkpoint and the layer-decomposition checkpoint use different encoder weights. Pick the version that matches the model you are running:
- Design - for Ming Image Design and anything finetuned from it. Available in BF16, int8 and w4a8.
- Layer - for Ming Image Design Layer. Available in BF16 and int8.
Loading the wrong one does not error cleanly; it produces noise. If you are unsure which you have, match the filename - the Layer weights carry layer in the name.
Pick one file, not all of them
Each file is the whole encoder at a different precision, so you need exactly one. BF16 is the reference weights at full size. int8 is the practical default for most setups. w4a8 (Design only) is the smallest and the one to reach for on limited VRAM. Every file is marked optional for that reason - nothing here requires anything else here.
Where these go
In ComfyUI, models/text_encoders/. This is a component rather than a runnable model, so there is no Generate button and no on-site generation - if you want to generate with Ming, that is the Ming Image Design page, where the encoder is already handled for you.
Links
- Source: Ming-Image-0.1-Design and Ming-Image-0.1-Design-Layer
- Repackaged files: Comfy-Org/Ming-Image
- Code: github.com/inclusionAI/Ming-Image