HiDream-O1-Image

Overview

HiDream-O1-Image is an image generation foundation model from HiDream.ai, released on May 8, 2026 under the MIT License. Its Hugging Face repository was created at 2026-05-08T13:06:04Z, and the model card's Project Updates log records: "May 8, 2026: We've open-sourced HiDream-O1-Image (8B), including both the undistilled and distilled Dev variants, together with the Reasoning-Driven Prompt Agent."

The architecture is the reason to care. HiDream describes it in one line — "One end-to-end model on raw pixels, no VAE, no disjoint text encoder" — and the arXiv abstract restates it: the Pixel-level Unified Transformer (UiT) works by "mapping raw image pixels, text tokens, and task-specific conditions into a single shared token space", which "eliminates the need for separate VAEs or disjoint pre-trained text encoders." Practically every rival — Stable Diffusion, FLUX, Qwen-Image, Z-Image — is a latent diffusion model: an autoencoder compresses the image roughly 8x per side, the transformer denoises in that latent space, and a decoder reconstructs pixels. HiDream deletes both halves of the autoencoder and one whole text encoder, and pushes the transformer down onto raw pixels.

The results are strong and the distribution is small. On Artificial Analysis's text-to-image arena, the closed HiDream-O1-Image-1.5 sits at #4 of 149 ranked models with an Elo of 1,265 — the highest-placed model from any Chinese developer, ahead of OpenAI's GPT Image 1.5 (high). The open, MIT-licensed HiDream-O1-Image-Dev-2604 is #2 among open-weights models. And yet HiDream-ai/HiDream-O1-Image has 22,788 Hugging Face downloads in the trailing 30 days, against 933,580 for Alibaba's Z-Image-Turbo — a model 8 arena places below it on that same board (#69 against #61). The #4 result belongs to the checkpoint you cannot download; the checkpoint you can download sits at #61. Downloads measure size, licence friction, and tooling support; they do not measure quality.

HiDream.ai (北京智象未来科技有限公司, Beijing Zhixiang Future Technology) is a Chinese startup founded in March 2023, whose stated mission is "to provide responsible intelligence for a better future." Its founder and CEO is Dr. Tao Mei (the corporate site renders his name Dr. Mei Tao). His prior roles at JD.com and Microsoft Research are widely reported elsewhere but appear nowhere on HiDream's own site, so this page does not assert them.

A note on naming

The HiDream O1 family has an open tier and a closed tier, and the names do not warn you which is which.

MIT, open weights, downloadable today:

Closed. No weights, anywhere:

• HiDream-O1-Image-1.5 — the current flagship. hidream.ai's hero headline reads "HiDream O1 Image 1.5" over the strapline "Natively unified. The latest flagship." The HiDream-ai Hugging Face organisation contains fourteen repositories and none is named 1.5; the API returns 401 for HiDream-ai/HiDream-O1-Image-1.5 (though a repo id that has never existed also returns 401, so that signal alone proves nothing). Artificial Analysis stores openWeightsUrl: null against its 1.5 row while linking real Hugging Face URLs for the 8B variants. AA's own note on the row: "HiDream-O1-Image-1.5 is the latest version of HiDream O1. Earlier variants, including HiDream-O1-Image, HiDream-O1-Image-Dev, and HiDream-O1-Image-Dev-2604, are benchmarked and listed separately."
• HiDream-O1-Image-Pro (200B+) — appears as the top row of every benchmark table in HiDream's own model card and technical report. No weights, no repository, no arena entry. HiDream publishes no statement equating Pro with 1.5, so this page does not equate them.

HiDream-O1-Image is not HiDream-I1. HiDream-I1-Full / -Dev / -Fast are the April 2025 latent-diffusion generation. They are also MIT, HiDream-I1-Fast is still the org's most-downloaded model (51,789 downloads/30d, more than twice the new flagship), and they are a different architecture entirely.

Capabilities

• One model, many tasks. Text-to-image, instruction-based image editing, subject-driven personalization from multiple references, long-text rendering, layout and skeleton conditioning, and storyboard generation — all from a single set of weights, with no task-specific adapters.
• No VAE, no text encoder. The UiT natively ingests raw pixels, text tokens and task conditions in one token stream. Nothing is compressed into a latent space and nothing is embedded by a frozen external encoder.
• Native 2,048 x 2,048 synthesis. HiDream calls this "Direct synthesis up to 2,048 × 2,048 with sharp fine-grained detail" — not an upscale pass.
• Reasoning-Driven Prompt Agent. A separate "thinking" agent (see chain-of-thought) that "explicitly reasons through layout, subject attributes, physical logic, and text-rendering details, then rewrites a raw user instruction into a self-contained English prompt." It is a distinct 31B model, not part of the 8B.
• Bilingual long-text rendering. HiDream reports 0.979 (EN) and 0.978 (ZH) on LongText-Bench for the 8B. In HiDream's own table the 8B places second on ZH but third on EN, behind Nano Banana 2.0 (0.980) as well as HiDream's unreleased Pro.
• Two sampling budgets. The full model runs 50 inference steps; the Dev variants run 28, via distillation.

Technical Specifications

• Hugging Face repositories: HiDream-ai/HiDream-O1-Image, HiDream-ai/HiDream-O1-Image-Dev, HiDream-ai/HiDream-O1-Image-Dev-2604
• Parameters: total_parameters: 8,804,887,792 (8.80B) per the checkpoint's model.safetensors.index.json — the whole model, encoder included. Dev-2604 reports an identical count.
• Architecture: Pixel-level Unified Transformer (UiT). The config.json declares "architectures": ["Qwen3VLForConditionalGeneration"] and "model_type": "qwen3_vl"; text tower 4,096 hidden / 36 layers / 32 attention heads / 8 KV heads, vision tower depth 27, patch size 16. HiDream's documentation does not discuss this lineage.
• VAE: none. This is the defining property of the model.
• Precision / disk: FP32 weights, total_size: 35,219,551,168 bytes (35.22 GB) across 8 safetensors shards
• Resolution: up to 2,048 x 2,048
• Sampling steps: 50 (full), 28 (Dev, Dev-2604)
• Prompt agent: google/gemma-4-31B-it for the base model; HiDream-ai/Prompt-Refine (a Gemma-4-31B-it finetune, MIT) for Dev-2604
• Attention kernel: flash-attn strongly recommended; without it you must edit models/pipeline.py line 341 or inference fails
• License: MIT — on all three repositories and on the GitHub repo
• Paper: arXiv:2605.11061, submitted May 11, 2026
• HiDream does not publish a VRAM requirement, a training compute figure, or a training-data description, so this page states none.

Use Cases

• Poster, packaging and signage design with heavy embedded text — HiDream self-reports 0.979 (EN) and 0.978 (ZH) on the LongText-Bench benchmark, within 0.001 of the best score in its own table.
• Editing and generation in one deployment — one checkpoint replaces a separate text-to-image model, an editing model, and an IP-adapter stack.
• Subject-driven personalization — preserving a character or product identity across new scenes from multiple reference images.
• Storyboarding and layout-conditioned generation — the May 13, 2026 update added layout and skeleton conditioning to the IP pipeline.
• Pixel-space generation research — the only competitive, permissively licensed, VAE-free image model to study. Latent-space artefacts simply do not exist here.
• Commercially unconstrained products — MIT has no revenue cap, unlike Stable Diffusion 3.5's $1M Community License threshold.

Performance / Benchmarks

arena.ai (the former LMArena) and Artificial Analysis are different organisations running different Elo scales. Their numbers are never comparable to each other.

Artificial Analysis text-to-image arena (fetched 2026-07-08; 149 ranked models, 133 marked current)

Three observations, all uncomfortable for someone:

1. HiDream-O1-Image-1.5 is the highest-ranked Chinese image model on this board, at #4, with a 69% win rate and a 95% confidence interval of 1,256–1,274. Ranks 1, 2, 3 and 5 are OpenAI, Reve, Microsoft AI and OpenAI. It is closed.
2. The open Dev-2604 outranks the open full model by 76 Elo — #20 versus #61. The 28-step distilled checkpoint paired with a prompt refiner beats the 50-step undistilled one. HiDream's model card does not explain this.
3. Dev-2604 is #2 open-weights, not #1. AA's own FAQ: "Cosmos3-Super-Text2Image (agentic) currently leads among open weights models in the Artificial Analysis Text to Image Arena with an Elo score of 1226, followed by HiDream-O1-Image-Dev-2604 (Elo 1187) and ERNIE Image (Elo 1168)."

The vendor's stale banner

HiDream's model card still leads with: "HiDream-O1-Image-Dev-2604 debuts at #8 in the Artificial Analysis Text to Image Arena, which is positioned to be the new leading open weights Text to Image model." That claim dates to mid-May 2026 and is no longer true. NVIDIA's Cosmos3-Super-Text2Image shipped 2026-05-31 (AA's recorded release date) and now holds both the #8 slot and the open-weights lead. Dev-2604 has slipped to #20.

arena.ai text-to-image leaderboard (fetched 2026-07-08; 72 models)

Only the full open checkpoint is ranked here, at #34 of 72. hidream-o1-image-1.5 is registered as a votable model in arena.ai's roster but carries no leaderboard rank, so arena.ai offers no independent read on the flagship.

Vendor-reported benchmarks (HiDream's own model card — treat as vendor claims)

In all five tables HiDream awards first place to its own unreleased Pro model, and in four of the five it awards second place to its own 8B — the exception is LongText-Bench-EN, where Nano Banana 2.0 (0.980) edges the 8B (0.979). HiDream summarises: "With only 8B parameters, achieves performance parity with or even surpasses larger open-source DiTs and leading closed-source models." These are self-run evaluations, published without third-party replication.

Limitations

• The flagship is closed. The #4 arena result belongs to HiDream-O1-Image-1.5, which you cannot download, inspect, or self-host. The MIT weights you can download rank #20 (Dev-2604) and #61 (full) on the same board.
• HiDream-O1-Image-Pro may not be a product at all. It is a 200B+ column in HiDream's own tables with no weights, no API listing, and no arena presence. Every "beats GPT Image 2" headline traceable to it is a claim about a model no one outside HiDream has used.
• "8B" excludes a 31B prompt agent. The Reasoning-Driven Prompt Agent runs google/gemma-4-31B-it, and Dev-2604's recommended pipeline runs HiDream-ai/Prompt-Refine, a Gemma-4-31B-it finetune. The leaderboard-winning configuration is roughly 40B of weights, not 8B.
• Every quality number in the model card is a vendor claim. HiDream ran all five benchmark suites itself and takes first place in every one, and second place in four of the five. Both independent arenas rank the open checkpoints far lower than the tables imply.
• The model card's headline claim is stale. "the new leading open weights Text to Image model" has been false since NVIDIA's Cosmos3-Super-Text2Image landed on 2026-05-31.
• 35.22 GB of FP32 weights, and no published VRAM figure. HiDream states no memory requirement, and ships no quantised release. flash-attn is effectively mandatory. There is no consumer-GPU story comparable to Z-Image's.
• Essentially no ecosystem. 1,203 GitHub stars against Z-Image's roughly 11,700; the diffusers pipeline PR (#13749) is still unmerged; no Hugging Face inference provider is mapped; 3,257 all-time downloads for Dev-2604. Inference requires HiDream's own repository or a hosted endpoint at WaveSpeed or fal.ai.
• Nothing is disclosed about training. No dataset description, no compute budget, no data-provenance statement.
• Last commit June 22, 2026 — and the only update that day was a Discord link.

Pricing & Access

Self-hosting

Free, under a real open-source licence. MIT on HiDream-O1-Image, HiDream-O1-Image-Dev, HiDream-O1-Image-Dev-2604, Prompt-Refine, and on the GitHub repository (confirmed via the GitHub API: "spdx_id": "MIT"). No revenue cap, no field-of-use restriction, no geographic exclusion. Training data is not released, so this is open weights plus an open licence — not open data.

Hosted access

HiDream publishes no rate card we could verify. hidream.ai displays no pricing. Hugging Face's inferenceProviderMapping is empty for both HiDream-O1-Image and HiDream-O1-Image-Dev-2604 — no HF-routed provider serves them.

Free demos run on Hugging Face Spaces: HiDream-ai/HiDream-O1-Image (139 likes), HiDream-ai/HiDream-O1-Image-Dev, HiDream-ai/HiDream-O1-Image-Dev-2604.

HiDream's consumer product

Vivago.ai is real and it is HiDream's. The HiDream-ai GitHub organisation lists https://vivago.ai/ as its website, hidreamai.com links out to vivago.ai, and Artificial Analysis benchmarks Vivago 2.1 (#47, Elo 1,135, released 2025-10-20) and Vivago 2.0 (#63, Elo 1,108, released 2025-06-10) with the creator field set to HiDream. AA also lists a vivago_hidream-i1-fast endpoint. HiDream's own product pages name Pixeling, PixMaker and HiHarness. We have verified none of Vivago's own product claims, so this page states nothing about its features, user numbers, pricing, or funding.

Ecosystem & Tools

• github.com/HiDream-ai/HiDream-O1-Image — HiDream's only supported self-hosting path. 1,203 stars, 33 forks, MIT, last pushed 2026-06-22. inference.py (main) for the full model, dev branch for Dev-2604.
• app.py — a single-file Flask web UI shipped in the repo, exposing every generation mode plus the Prompt Agent, on port 7860.
• HiDream-ai/Prompt-Refine — the Gemma-4-31B-it finetune that Dev-2604's prompt-engineering pipeline depends on. MIT.
• Hugging Face Spaces — official hosted demos for all three open checkpoints.
• WaveSpeed and fal.ai — the two third-party hosts Artificial Analysis lists for the O1 generation, each serving HiDream-O1-Image and HiDream-O1-Image-Dev. AA lists no endpoint for Dev-2604 or 1.5.
• diffusers PR #13749 — "feat: Add HiDream-O1 transformer and image generation pipeline", still open as of July 8, 2026. Until it merges there is no diffusers path; the checkpoint loads through transformers plus HiDream's own models/pipeline.py.
• ComfyUI, ModelScope — linked as ecosystem targets from hidream.ai.
• Discord — opened June 22, 2026.

Community & Resources

• HiDream-O1-Image: A Natively Unified Image Generative Foundation Model with Pixel-level Unified Transformer — the technical report (submitted 2026-05-11); source of the UiT and no-VAE claims
• HiDream-ai/HiDream-O1-Image on Hugging Face — 8.80B FP32 weights, MIT, the Project Updates log, all five benchmark tables
• HiDream-ai/HiDream-O1-Image-Dev-2604 on Hugging Face — the 28-step checkpoint that AA ranks #2 open-weights
• HiDream-ai/HiDream-O1-Image on GitHub — code, MIT LICENSE, dev branch
• hidream.ai — the project landing page, where HiDream O1 Image 1.5 is called "the latest flagship"
• hidreamai.com — the corporate site: 北京智象未来科技有限公司, founded March 2023
• Artificial Analysis text-to-image arena — where 1.5 is #4 and Dev-2604 is #20
• arena.ai text-to-image leaderboard — where hidream-o1-image is #34 of 72
• Compare with Z-Image, Qwen-Image 2.0, HunyuanImage 3.0, Seedream 5.0, and Stable Diffusion 3.5