alexokita/Ming-Image-0.1-Design-FP8

Ming-Image-0.1-Design (FP8 E4M3)

Weight-only FP8 E4M3 derivative of inclusionAI/Ming-Image-0.1-Design.

This is not an inclusionAI release. The original weights are MIT licensed; this repository redistributes a quantized copy under that same license. Copyright in the model remains with inclusionAI.

What was quantized

Linear weights in the thinker (mllm/), the Qwen2 connector (connector/), and the diffusion transformer (transformer/) are stored with per-output-channel absmax into float8_e4m3fn. A matrix is quantized only when it has at least 1,048,576 parameters. For int4, the input dimension must also be divisible by 128; other matrices stay bf16.

Each quantized tensor keeps its original name. The matching scale is a float32 tensor named by replacing the .weight suffix with .weight_scale.

These pieces are copied from the original checkpoint unchanged, in bf16:

• token embeddings and the unused lm_head
• layer norms and biases
• the MLP condition projector (mlp/)
• the 4-channel RGBA VAE (vae/)
• the scheduler and tokenizer

The method is data-free absmax. No calibration prompts were run, so this is not GPTQ, AWQ, or SmoothQuant.

FP8 E4M3 storage is about half the bf16 size of each quantized matrix. Activations are still computed in bf16.

How to use

The ComfyUI node Ming Image 0.1 Design loads this folder when its weights setting is fp8. Put the repository here, relative to the ComfyUI root:

models/ming-image/Ming-Image-0.1-Design-FP8

The node does not download quantized checkpoints. From a shell:

hf download alexokita/Ming-Image-0.1-Design-FP8 --local-dir <ComfyUI>/models/ming-image/Ming-Image-0.1-Design-FP8

The folder must contain quantization_config.json. Sampling settings from the original model card are 12 steps, CFG 1.0, and either 1024 or 2048 resolution. The workflow is https://huggingface.co/alexokita/Ming-Image-0.1-Design-ComfyUI.

The original model card, architecture notes, and license are here: https://huggingface.co/inclusionAI/Ming-Image-0.1-Design

Limitations

• Int4 in particular can soften small type and fine UI detail. fp8 is the closest of the three to the bf16 checkpoint.
• Transparent-background trigger phrases from the original model card were not published with the weights. The VAE is still the original 4-channel decoder, so an alpha channel is produced either way.
• This file format is for the Ming-Image ComfyUI node in this project. It is not a Diffusers model_index.json pipeline.