voltaire321/Ming-Image-0.1-Design-GGUF

Ming-Image 0.1 Design — GGUF for stable-diffusion.cpp

GGUF conversions of Ming-Image-0.1-Design by inclusionAI, for running with stable-diffusion.cpp. Quantized only — no fine-tuning or merging; all credit to inclusionAI.

Published for Haruspex, whose bundled image engine downloads them. On Vulkan without bf16 support (AMD RADV, for example), the int8 and bf16 safetensors run on slow paths — about 120 s a step — while these GGUFs make a 1024×1024 image in about 35 s end to end.

Files

File	What	Type	Size	SHA-256
ming_image_0.1_design-Q8_0.gguf	diffusion model (DiT, 6B)	Q8_0	6.54 GB	260ded24503b35b80b6eb4a5949f664b26f6c00b46e8ec6e6d00942593eb120b
ming_ling_mini_2.0-Q4_K.gguf	text encoder (Ling-mini-2.0)	Q4_K	10.52 GB	70f7ea31d3e1ffa917e70b961e5d422b2a8b19327c3522033a381b1a4b766002

Also needed, unchanged from upstream:

• VAE: vae/ming_image_vae_bf16.safetensors (Comfy-Org)
• Tokenizer: mllm/tokenizer.json (inclusionAI)

A Q4_K diffusion model was tried and is not published: it lost the model's pixel-art rendering (smooth, painted output at the same seed).

How they were made

From Comfy-Org's bf16 safetensors (diffusion_models/ming_image_0.1_design_bf16.safetensors, text_encoders/ming_image_0.1_ling_mini_2.0_bf16.safetensors), with stable-diffusion.cpp master-929-3f8527a:

sd-cli -M convert --diffusion-model ming_image_0.1_design_bf16.safetensors \
  --type q8_0 -o ming_image_0.1_design-Q8_0.gguf
sd-cli -M convert -m ming_image_0.1_ling_mini_2.0_bf16.safetensors \
  --type q4_K -o ming_ling_mini_2.0-Q4_K.gguf

Use

sd-cli --diffusion-model ming_image_0.1_design-Q8_0.gguf \
  --llm ming_ling_mini_2.0-Q4_K.gguf \
  --vae ming_image_vae_bf16.safetensors --tokenizer tokenizer.json \
  -p "A cheerful orange cat sticker" \
  -W 1024 -H 1024 --steps 12 --cfg-scale 1 --sampling-method euler --diffusion-fa \
  -o out.png

Transparent output. Ming-Image's RGBA prompt phrases alone did not give alpha in our tests (also reported upstream, inclusionAI/Ming-Image#5). What works: start from a fully transparent canvas at strength 0.9 and keep an RGBA phrase in the prompt — neither alone is enough:

# clear.png: a 1024×1024 RGBA image with alpha 0 everywhere
sd-cli ... -i clear.png --strength 0.9 \
  -p "RGBA, 4-channel, transparent background. A gold coin, pixel art"

License

MIT, as the originals: Ming-Image-0.1-Design and Ling-mini-2.0 are both released by inclusionAI under the MIT License (see LICENSE, and Ling-V2's LICENCE).