inclusionAI

Shipping Ming-Image Design: size the card by its files, and check the alpha in code Say you want an image endpoint for app mockups, slides and stickers, and Ming-Image-0.1-Design from inclusionAI looks like the right fit. It is MIT licensed and trained for layouts full of text. Its PNGs always carry an alpha channel, a fourth layer that stores the opacity of every pixel. Before you ship it, three things decide whether it behaves: which GPU you rent, how you phrase the transparency request, and what happens when a result is too big to send back. We ran it on one NVIDIA A100 80GB on our cloud and wrote up every number in the Ming-Image Design test on one A100. Budget memory from the component list The model card calls it a 6B model. That figure covers only the diffusion transformer, the part that refines random noise into an image step by step, and it is 12.3 GB on disk. The repository also ships a 34 GB language model that interprets the prompt, a 6.2 GB connector between the two and a 0.25 GB decoder that produces the pixels: 52.9 GB in total. Once loaded, the model held 46.3 GiB of GPU memory before drawing anything. A 1024 x 1024 image peaked at 50.45 GiB, and a 2048 x 2048 one reached 64.4 GiB. That lines up with the authors' GitHub README, where the smallest setup they validated is one GPU with at least 80 GiB. If your capacity planning starts from a parameter count, add up the folder sizes first. We also planned to fake a 24 GB card by capping the A100's memory. An open bug kept the new setting from reaching the running model, so nothing under 80 GB is tested. Treat transparency as a prompt format you have to test inclusionAI lists ten trigger phrases for transparent output and asks you to open the prompt with a single one. There is no special code path: the decoder always emits four channels, and transparency happens only if the model draws it. We wrote the request six different ways, and a single version produced a cutout. It started with "RGBA, 4-channel, transparent background" as a line by itself, followed by the fox sticker description: 83.2 % of pixels came out fully transparent. Put a comma after the phrase instead and the fox sat on opaque white, with a minimum alpha of 247. Yet the same newline trick on an app icon still gave a solid background. The newline helped the fox, and it guarantees nothing. The practical rule is to stop trusting the request and measure the response: prompt = "RGBA, 4-channel, transparent background\n" + description # newline, not ", " alpha = np.asarray(Image.open("out.png").getchannel("A")) transparent = (alpha == 0).mean() # fox with newline: 0.832, with a comma: 0.0 Retry or reject below a threshold that fits your assets. Our backend (the small server that keeps the model in memory and serves requests) returns this fraction with every image. Quoted strings are the spec, the rest is filler In 5 of our 7 layouts, every string we put in quotes was correct. Anything unspecified, such as chart axes or feature cards, filled up with pseudo-words that pass at a glance. !Menu titled TASCA DO RIO: Bifana reads 6.0l) instead of 6.00, Sardinhas 1110.0 instead of 11.00, plus an invented Bica 0.50 line Long lists slip. A 12-line menu asked for in two columns came out in one, with two wrong prices and an invented line. If your output carries prices, someone has to read every line. inclusionAI suggests first rewriting prompts into a structured, Figma-style description with another language model; we did not try that. Three settings that bite between the model and your client | What | Why it matters | |---|---| | 12 steps, guidance 1.0 | Steps are drawing passes; guidance is how hard the prompt steers them. The authors validated these. Our panel's image tester defaults to 20 and 4.5. | | 1024 and 2048 only | It draws at the closer level, then resizes: a 1920 x 1080 slide is drawn at 1024. | | Result size | Our test setup accepted answers up to 8 MB. Two busy 2048 images went over, got discarded, and the client sat until a 300-second timeout. | Resolution alone does not predict file size: other 2048 images fit. The SDK example's backend now lowers WebP quality until the file fits, tested on saved images only. The long platform wait is still open, and we did not check the hosted limit. Speed with the authors' untuned script: 6.2 to 7.4 seconds of GPU time per 1024 image over 10 runs, 35.8 seconds per 2048 image. The vLLM-Omni serving engine reports about 2.2 seconds at 1024 on two H100s; we did not run it. What this does not prove These are observations, not a quality score: 15 images, one try per prompt, prompts we wrote, one seed (the random starting point), a single GPU model, and requests sent one by one. We did not call the hosted service. On Python 3.11 the authors' pinned versions did not run as written; slightly newer numpy and torch fixed it. The dependency fixes, per-size memory figures and the full transparency table sit in the engineering appendix. To compare text rendering, our MOSS-Image hands-on test measured the same kind of layouts on one A40. Credit where it belongs. inclusionAI made Ming-Image-0.1-Design and released it under MIT, inference code and transparency phrases included. It reuses pieces from other teams: Ling for the language model, Qwen2.5 for vision, Qwen2 for the connector, Z-Image for the transformer design and Qwen-Image for the decoder. The vLLM-Omni team contributed a serving recipe plus benchmark numbers, hugging-apps a demo Space, and the AI Search channel is where we first saw the model. We wrote the custom backend, deployed it, fixed the dependencies and took the measurements. Timings, sample images and the deploy walkthrough live in the Ming-Image Design research article on inferencekey.com.

IDM (inclusionAI): a better DMD distill for 1–2 step text-to-image, matching stats of the model's own EMA features. On Z-Image it renders long text in 2 steps, beating 8-step Z-Image-Turbo on Chinese text. Weights coming. inclusionAI/IDM arxiv.org t.co

Paper: paperswithcode.co Project: inclusionai.github.io Code: inclusionAI/IDM

Intrinsic Distribution Matching One-step and few-step text-to-image generation must accurately render the details specified in the prompt despite limited opportunities for refinement. Distribution matching distillation (DMD) uses score differences to bring the generated distribution closer to a reference distribution. Following prior work, we estimate this difference with internal velocity branches, avoiding a separate fake-score network. We ask whether an explicit moment constraint improves this learned score supervision. A decomposition of the reverse-KL field into a Gaussian moment component and a non-Gaussian residual identifies statistics that can be constrained explicitly. We introduce Intrinsic Distribution Matching (IDM), which retains the full DMD correction and matches feature means and per-coordinate standard deviations using the generator's exponential-moving-average (EMA) copy, requiring no external feature encoder. Experiments use full-parameter and low-rank adaptation (LoRA) tuning on Z-Image-Turbo 6B and Qwen-Image 20B. At two network evaluations, IDM reaches 94.25% and 90.20% on Chinese and English LongText-Bench, the highest scores among the compared methods at this budget. Controlled ablations show gains from the added moment loss, while generator EMA features perform competitively with an external DINOv2 encoder across the five reported metrics. Code is available at inclusionAI/IDM.

Explore the full technical work: NVIDIA technical blog:developer.nvidia.com Ling-3.0-flash on Hugging Face:huggingface.co Ling-3.0-flash on ModelScope: modelscope.ai Deployment cookbook:inclusionAI/ling-cookbook Join the Ling community on Discord:discord.gg

Learning to Look Beyond Screens: Transferable Visual Observation from GUI Grounding Visual perception is evolving from passive recognition toward active, task-oriented observation. We investigate GUI grounding as a scalable source of spatial supervision for learning observation behaviors that transfer beyond interfaces. We introduce LBS (Look Beyond Screens), a vision-language model trained through supervised initialization on teacher-generated GUI grounding trajectories, followed by reinforcement learning for multi-turn crop-and-inspect decisions. Under fixed inference settings, LBS models based on Qwen3.5-35B-A3B and Qwen3.8-27B consistently improve over their corresponding baselines. Gains beyond supervised fine-tuning transfer to PerceptionBench for both backbones and to HR-Bench and ChartQAPro for Qwen3.5-35B-A3B, supporting observation strategies that generalize beyond GUI-specific imitation. Further analysis suggests complementary roles for the two stages: SFT establishes visual inspection primitives, while RL improves their downstream effectiveness. These findings support GUI grounding as a practical supervision paradigm for transferable active visual perception and broader multimodal reasoning. Our code and model are publicly available at github.com/inclusionAI/UI-Venus/tree/LBS and huggingface.co, respectively. Comments: Technical Report

A free 0.5B voice model now scores 1073 in blind listening tests. ElevenLabs v3 scores 1074. It's Ming 3 TTS from Ant Group, MIT licensed. ElevenLabs v4 is still well ahead, to be fair. For a side project, I'd stop paying per character. inclusionAI/Ming-3.0-tts t.co

10 new repos this week vs $2,335 a year of subscriptions Ten open-source projects, all created in the last nine days, against $2,335 a year of subscriptions — every price read at the vendor today. 00:00 Intro 00:30 Number 1 — OpenTRAP 01:24 Number 2 — AutoIllustrate 02:19 Number 3 — FlashVSR+ 03:14 Number 4 — Ming-3.0-tts 04:13 Number 5 — editonny 05:10 Number 6 — codex-openpencil 06:01 Number 7 — Indra OCR 06:58 Number 8 — SpeakMore 07:56 Number 9 — Morrow 08:45 Number 10 — ElixirOffice 09:44 Still worth paying for The ten: 1. OpenTRAP — a web-application security scanner for your own machine, built on the OWASP ZAP engine; not an AI tool, and the video says so, instead of Burp Suite Professional ($499, basis unlabelled on the page): opentrap/opentrap 2. AutoIllustrate — a Sakana AI agent that turns a written description into a scientific figure, instead of BioRender Academic Individual ($420/yr at $35/mo billed annually): SakanaAI/AutoIllustrate 3. FlashVSR+ — video super-resolution research code that upscales footage on an NVIDIA GPU, instead of Topaz Video Personal ($299/yr as the page states): OpenImagingLab/FlashVSR_Plus 4. Ming-3.0-tts — an open text-to-speech model family in a small dense and a larger mixture-of-experts version, instead of ElevenLabs Creator ($219.96/yr at the page's annual $18.33): inclusionAI/Ming-3.0-tts 5. editonny — a video-editing skill that has Claude Code or Codex do the cutting, instead of Descript Hobbyist ($192/yr at the lower $16 the page prints): soumatheusgomes/editonny 6. codex-openpencil — a Codex plugin that opens an interactive design canvas inside your coding agent, instead of Figma Professional ($192/yr at the $16 full seat): ZSeven-W/codex-openpencil 7. Indra OCR — a small OCR model that reads documents on a CPU, instead of Amazon Textract Detect Document Text ($150/yr at 100,000 pages and $0.0015 a page): datalab-to/indra 8. SpeakMore — desktop dictation that runs through whichever speech model you point it at, instead of Typeless Pro ($144/yr on the yearly $12 seat): zephyr4123/speakmore 9. Morrow — a local notes and workspace app that keeps every page as plain Markdown; no AI of its own, and the video says so, instead of Notion Plus ($120/yr at the $10 the page prints): Aarav2709/Morrow 10. ElixirOffice — a desktop office suite for people used to Microsoft Office, with AI optional and off until you configure it, instead of Microsoft 365 Personal ($99.99/yr): leag1234/elixiroffice Seven of the ten are AI tools, one has AI as an option you switch on, and two — OpenTRAP and Morrow — are not AI tools at all. The video names them. Six of the ten pairings are mine rather than the project's own claim, and the narration says which. Four vendor pages print a price with no annual or monthly label — Burp Suite, Descript, Figma and Notion — and in every case the lower figure was used and the video says so instead of guessing. Amazon Textract charges per page, not per month, so the $150 is an assumption of 100,000 pages a year at $0.0015 a page; halve the volume and you halve the saving. Three of the ten only work on top of a coding agent you already pay for, and that subscription is not included in any figure here. Every tool's catch is in the video. If one of these saved you a subscription, like this video and subscribe. New AI tools, tested and priced, every episode. #AItools #opensource #selfhosted #FUBARAI

GeekPie98 - Rambo/ComfyYouEye-Ming-3-tts: Native Comfy You Eye port of Ming-3.0-tts, supporting 0.5B and Tiny 9B speech generation. ComfyYouEye Ming 3.0 TTS Ming-3.0-tts 的原生 ComfyYouEye 移植，支持 0.5B 和 Tiny 9B 语音生成。 Saganaki22/ComfyUI-Ming-3-tts THANKS Rambo.

Ming-3.0-tts: open weights, MIT. #1 overall on Artificial Analysis Pronunciation Robustness. LM → DiT flow head → continuous 12.5Hz latents. No discrete tokens. 44.1kHz. 0.5B dense + 7.9B MoE (1.3B active). 7 languages. 14 paralinguistic tags. inclusionAI/Ming-3.0-tts