Ming-Image GGUFs v1.0 - Workflows for Other

GGUF quantizations of Ming-Image-0.1-Design for use with ComfyUI-GGUF.

GGUFs:
https://huggingface.co/realrebelai/Ming-Image_GGUFs

REQUIRES PR for GGUF NODES:

https://github.com/city96/ComfyUI-GGUF/pull/484

Original model:
https://huggingface.co/inclusionAI/Ming-Image-0.1-Design

Official Ming-Image repository:
https://github.com/inclusionAI/Ming-Image

These are not simple one-pass quants. The diffusion-model files use a mixed-precision HQ ladder designed to keep the most sensitive Ming-Image tensors at higher precision while aggressively quantizing the larger bulk weights.

Diffusion model quantization

The diffusion model contains 519 tensors.

The process used for the files in this repository was:

Convert the original BF16 diffusion weights into a shape-preserving BF16 GGUF master.
Quantize that BF16 GGUF directly with llama-quantize to create donor files:
- Q8_0
- Q6_K
- Q5_K_M
- Q4_K_M
- Q3_K_M
- Q2_K
Build the final -HQ models by selecting tensors from those already-quantized donor files.
The HQ assembly step copies the existing packed GGUF tensor data directly. It does not dequantize and requantize each selected tensor.

The general.architecture transport metadata used during GGUF conversion is for quantizer/loader compatibility only. It should not be interpreted as Ming-Image being a Lumina2 model.

HQ mixed-precision policy

The same tensor groups are protected across the entire ladder.

Always kept in BF16

The following groups remain BF16 in every HQ quant:

t_embedder.*
all_x_embedder.*
all_final_layer.*
cap_embedder.*
context_refiner.*
noise_refiner.*
*.adaLN_modulation.*

This includes the input/output projections, time conditioning, both refiner stacks, caption projection path, and all adaLN modulation tensors.

There are 309 BF16 tensors retained in the HQ models.

Attention tensors

The following attention weights are assigned a higher precision than the bulk quant where appropriate:

*.attention.to_q.weight
*.attention.to_k.weight
*.attention.to_v.weight
*.attention.to_out.0.weight

There are 120 attention tensors in this protected group.

FFN down projection

The FFN down projection is also protected separately:

*.feed_forward.w2.weight

There are 30 protected FFN-down tensors.

Bulk tensors

The remaining 60 tensors use the base quantization level of that rung.

This includes the large FFN w1 / w3 tensors unless otherwise covered by one of the protected groups above.

Exact diffusion HQ ladder

File / rungBF16 critical tensorsAttention Q/K/V/OFFN w2Remaining bulkQ8_0-HQBF16Q8_0Q8_0Q8_0Q6_K-HQBF16Q8_0Q8_0Q6_KQ5_K_M-HQBF16Q8_0Q6_KQ5_K_MQ4_K_M-HQBF16Q8_0Q6_KQ4_K_MQ3_K_M-HQBF16Q6_KQ5_K_MQ3_K_MQ2_K-HQBF16Q5_K_MQ4_K_MQ2_K

For example, the Q4_K_M-HQ model is composed of:

309 tensors  BF16
120 tensors  Q8_0
 30 tensors  Q6_K
 60 tensors  Q4_K
-------------------
519 tensors total

This is why the HQ files may be larger than a conventional uniform quant at the same advertised rung.

The goal is to preserve Ming-Image's conditioning, attention, refinement, and output quality while still reducing the memory footprint of the largest weight groups.

Ling Mini 2.0 text encoder GGUFs

This repository also includes low-VRAM GGUF variants of the Ming-Image Ling Mini 2.0 text encoder.

The source encoder contains approximately 17.3B parameters across 642 tensors, including large routed MoE expert banks.

Two practical variants were built:

Q4_K_M-HQ encoder

Small / critical routing and bridge tensors remain at high precision.
The rest of the encoder is primarily Q4_K_M.
Intended as the higher-quality low-VRAM encoder option.

Q2_K-HQ encoder

The Q2 encoder is intentionally not a uniform Q2 conversion.

Small / critical routing and bridge tensors remain at high precision.
The non-MoE backbone remains at Q4_K_M.
The large routed MoE expert banks use Q2_K.

This concentrates the most aggressive compression on the enormous expert weights instead of applying Q2 indiscriminately to the whole encoder.

The encoder contains 38 large 3D expert tensors across the routed MoE layers, so this split makes a substantial difference to total size while protecting the shared backbone.

ComfyUI installation

Install:

https://github.com/city96/ComfyUI-GGUF

Place diffusion GGUF files in:

ComfyUI/models/diffusion_models/

or the GGUF UNet/diffusion-model folder used by your ComfyUI installation.

Place Ling Mini GGUF text encoders in:

ComfyUI/models/text_encoders/

Place the Ming-Image VAE in:

ComfyUI/models/vae/

Use:

Unet Loader (GGUF) for the diffusion model
CLIP Loader (GGUF) for the Ling Mini encoder
the standard VAE Loader for the Ming-Image VAE

Important: Ming-Image GGUF loader support

Ming-Image stores its attention Q/K/V weights separately.

With quantized GGUF tensors, the physical packed byte width can differ from the tensor's logical width. For example, a logical width of 3840 can occupy 4080 packed bytes in Q8_0.

The normal ComfyUI Diffusers conversion path tries to merge these tensors with ordinary Torch assignment and can fail with an error like:

The expanded size of the tensor (3840) must match the existing size (4080)
Target sizes: [3840, 3840]
Tensor sizes: [3840, 4080]

A Ming-Image compatibility fix has been submitted for ComfyUI-GGUF that performs the Q/K/V merge while preserving the packed GGUF storage and the correct logical tensor shape.

Until that support is merged into ComfyUI-GGUF main, use the Ming-Image PR/branch for loading these diffusion GGUFs.

Add the ComfyUI-GGUF PR link here once published.

The GGUF files themselves do not need to be repaired, rebuilt, dequantized, or requantized.

Recommended starting point

For most low-VRAM users:

Diffusion: Ming-Image-0.1-Design-Q4_K_M-HQ.gguf
Encoder:   Ming-Image-0.1-Ling-Mini-2.0-Q4_K_M-HQ.gguf

If memory is extremely limited, move down to the Q3/Q2 diffusion files and/or the Q2 Ling Mini encoder.

If you have more memory and want to preserve more precision, move upward through Q5, Q6, or Q8.

Suggested generation settings

The original Ming-Image release recommends:

Resolution: 2048x2048
Steps:      12
CFG:        1.0

1024x1024 is also useful for faster testing.

See the official model card for current upstream recommendations:

https://huggingface.co/inclusionAI/Ming-Image-0.1-Design

Notes

These files are intended for ComfyUI-GGUF.
The -HQ suffix means the file uses the mixed-precision tensor policy documented above.
Quantization is lossy. Lower-bit variants trade some model fidelity for reduced memory/storage requirements.
The original model, architecture, license, and upstream documentation belong to the Ming-Image authors.
This repository provides GGUF conversions/quantizations only.

Credits

Ming-Image-0.1-Design:
https://huggingface.co/inclusionAI/Ming-Image-0.1-Design

Official Ming-Image GitHub:
https://github.com/inclusionAI/Ming-Image

ComfyUI-GGUF:
https://github.com/city96/ComfyUI-GGUF