Ming-Image 0.1 Design

Interactive text-to-image demo for inclusionAI/Ming-Image-0.1-Design — a 6B design-oriented diffusion model that produces complete, text-rich compositions (UI screens, infographics, posters, dashboards) and decodes to RGBA, so transparent backgrounds survive the VAE round-trip.

Recommended sampling settings from the model card: 12 steps, CFG 1.0, BF16, on one of the two supported resolution buckets — 1024 x 1024 (fast) or 2048 x 2048 (the configuration the authors validated on an 80 GB GPU).

Pipeline

The checkpoint ships no model_index.json and the inference repository the model card links to (github.com/inclusionAI/Ming-Image) is not public. The pipeline in ming/ follows the reference integration for this checkpoint, vLLM-Omni PR #8021 (pinned to a62d2ec999ae8fa669e8c22194c67575c0f2d3dc):

1. A BailingMoE-v2 "thinker" (20 layers, 256 experts) reads the raw prompt followed by 256 learned query tokens, with plain sequential positions for every token.
2. Query stream: the final hidden states at the 256 queries go through proj_in, a bidirectional Qwen2 connector and proj_out (width 2560).
3. Direct stream: the prompt tokens' own hidden states - before blocks 5 and 12 and after the final norm - are concatenated and projected by proj_directvlm (RMSNorm + Linear, width 3840). This stream carries the exact text to render; the checkpoint's use_vlm_directvlm_condition is true.
4. The Z-Image flow-matching DiT (dim 3840, 30 layers, 16-channel latents) takes the embedded query stream followed by the direct stream as its caption, with alignment padding zeroed and masked out of attention. Sampling uses a FlowMatchEulerDiscreteScheduler (dynamic resolution-dependent mu, sigma_min = 0) and noise_pred = -model_out.
5. CFG applies whenever the scale is above 0 - including the recommended 1.0 - as positive + scale * (positive - negative) against zero conditioning on both streams.
6. Decoding uses AutoencoderKLQwenImage with a 4-channel (RGBA) input/output.

Runtime

ZeroGPU xlarge (~96 GB): the full checkpoint is ~53 GB in bf16 (thinker ~33 GB, connector ~6 GB, DiT ~12 GB, VAE + MLP < 1 GB), and everything is loaded at module scope and moved to CUDA eagerly.

Assets

assets/showcase.webp and assets/transparency_showcase.webp are the sample images published in the model card (MIT-licensed, from inclusionAI/Ming-Image-0.1-Design).

Licence

Model weights are MIT licensed (see the model card). The Space code is provided for demonstration purposes.