--- license: mit base_model: inclusionAI/Ming-Image-0.1-Design pipeline_tag: text-to-image tags: - ming-image - gguf - comfyui - text-to-image - image-generation - quantization - low-vram --- # Ming-Image-0.1-Design GGUF
GGUF quantizations of **Ming-Image-0.1-Design** for use with **ComfyUI-GGUF**. ## UPDATE: Encoders have been updated as they were stripped of editing components before. **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: 1. Convert the original BF16 diffusion weights into a **shape-preserving BF16 GGUF master**. 2. 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` 3. Build the final `-HQ` models by selecting tensors from those already-quantized donor files. 4. 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**: ```text 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: ```text *.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: ```text *.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 / rung | BF16 critical tensors | Attention Q/K/V/O | FFN `w2` | Remaining bulk | |---|---:|---:|---:|---:| | **Q8_0-HQ** | BF16 | Q8_0 | Q8_0 | Q8_0 | | **Q6_K-HQ** | BF16 | Q8_0 | Q8_0 | Q6_K | | **Q5_K_M-HQ** | BF16 | Q8_0 | Q6_K | Q5_K_M | | **Q4_K_M-HQ** | BF16 | Q8_0 | Q6_K | Q4_K_M | | **Q3_K_M-HQ** | BF16 | Q6_K | Q5_K_M | Q3_K_M | | **Q2_K-HQ** | BF16 | Q5_K_M | Q4_K_M | Q2_K | For example, the **Q4_K_M-HQ** model is composed of: ```text 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: ```text ComfyUI/models/diffusion_models/ ``` or the GGUF UNet/diffusion-model folder used by your ComfyUI installation. Place Ling Mini GGUF text encoders in: ```text ComfyUI/models/text_encoders/ ``` Place the Ming-Image VAE in: ```text 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: ```text 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: ```text 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: ```text 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