--- title: "HiDream.ai's HiDream-O1-Image-1.5 Beats Google Gemini in Blind Image Votes" subtitle: "HiDream-O1-Image-1.5 lands at #3 on the Artificial Analysis leaderboard, beating Google's Gemini image model with a pixel-native architecture that ditches the VAE entirely" url: "https://alphasignal.ai/news/hidream-ai-s-hidream-o1-image-1-5-beats-google-gemini-in-blind-image-votes" author: "AlphaSignal Newsroom" author_url: "https://alphasignal.ai/editorial-team" company: "Artificial Analysis" company_url: "https://alphasignal.ai/company/artificialanlys" published: "2026-06-09T21:50:44.000Z" updated: "2026-06-10T07:28:08.261Z" type: "News" topics: - "Image" - "Benchmarks" subtopics: - "Image Generation" upvotes: 220 read_time: "6 min" source_url: "https://artificialanalysis.ai/image/leaderboard/text-to-image" comments: 0 --- # HiDream.ai's HiDream-O1-Image-1.5 Beats Google Gemini in Blind Image Votes > HiDream-O1-Image-1.5 lands at #3 on the Artificial Analysis leaderboard, beating Google's Gemini image model with a pixel-native architecture that ditches the VAE entirely The image generation leaderboard just got a new serious contender. [HiDream-O1-Image-1.5](https://artificialanalysis.ai/image/leaderboard/text-to-image), the latest closed-source model from HiDream.ai, has debuted at **#3** on the Artificial Analysis Text-to-Image Arena, surpassing Google's Nano Banana 2 (Gemini 3.1 Flash Image Preview) and sitting just behind OpenAI's GPT Image models. This is a big deal for a company that only open-sourced its first O1-series model a month ago. ## A quick recap: what is the O1 Image family? HiDream-O1-Image is a natively unified image generative foundation model built on a Pixel-level Unified Transformer (UiT) without external VAEs or disjoint text encoders, which natively encodes raw pixels, text, and task-specific conditions in a single shared token space, supporting text-to-image, image editing, and subject-driven personalization at up to 2,048 × 2,048. The original open-source 8B model launched and immediately became the highest-ranked open-weight model on the same leaderboard. The 1.5 release is a closed-source, more capable follow-up in the same family. The defining architectural choice is the elimination of the variational autoencoder (VAE). Every mainstream diffusion model , FLUX.2, Stable Diffusion 3.5, DALL-E 3 , compresses images through a VAE before and after the diffusion process. HiDream skips that step completely. The Unified Transformer maps raw image pixels, text tokens, and task-specific conditions into a single shared token space, removing the VAE bottleneck used by virtually every competing model. ## Why dropping the VAE matters Most image generation pipelines work in two stages: a text encoder converts your prompt into a vector, and a separate diffusion model generates an image in a compressed "latent space," which is then decoded back to pixels via the VAE. Each handoff introduces a translation bottleneck. Unlike latent DiTs that use latent-space VAE compression and pixel-space DiTs that typically rely on disjoint text encoders, the Unified Transformer in HiDream-O1-Image natively encodes raw image pixels, texts, and task-specific conditions within a shared token space, and thus generalizes to broader and more complex generative tasks. The practical payoff shows up most clearly in text rendering. The text-rendering numbers are the most interesting , CVTG-2K and LongText-Bench specifically stress the failure mode where latent-space models historically collapse, and HiDream-O1's pixel-native design is exactly the kind of change that should help there. HiDream O1 Image achieves 0.979 (English) and 0.978 (Chinese) on LongText-Bench, handling up to 5 distinct text regions within a single composition. For comparison, FLUX.2 [Dev] scores 0.963 EN and only 0.757 ZH on the same benchmark. ## How it was built The architecture is initialized from Qwen3-VL-8B-Instruct to leverage its multimodal pre-alignment, then trained through a three-stage progressive strategy. Stage I aligns pixel patches with language concepts at 512×512. Stage II expands to 1024×1024 and adds editing and personalization tasks. Stage III refines fine-grained detail at full 2048×2048 resolution. Post-training applies supervised fine-tuning followed by RLHF using GRPO, with reward signals covering OCR accuracy, aesthetics, instruction-following, and reasoning quality. The model also ships with a **Reasoning-Driven Prompt Agent** , a separate wrapper that runs on top of the generation model. HiDream-O1-Image ships with a Reasoning-Driven Prompt Agent 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. You can select either the OpenAI-compatible API backend (any endpoint, key, and model name) or the Local Gemma backend. ## What the benchmarks say HiDream-O1 wins every reported benchmark while being 7× smaller than FLUX.2 Dev and 3.4× smaller than Qwen-Image. That efficiency gap is the real headline from the technical paper. On GenEval, the 8B model scores 0.90 overall , beating GPT Image 1 (high) at 0.84, FLUX.2 [Dev] at 0.87, and Qwen-Image at 0.87. On HPSv3, a human preference benchmark, it scores 10.37 out of 12, ahead of GPT Image 2 at 10.21 and Nano Banana 2.0 at 10.01. The architecture also scales. The experimental HiDream-O1-Image-Pro proves this architecture scales up to 200B parameters without bottlenecking. The 200B+ Pro variant consistently pushes every benchmark number higher, suggesting the pixel-native approach doesn't hit a ceiling the way latent-space models do. ## What it's good at , and where it struggles The model's strongest differentiators: - **Text rendering** , multilingual, multi-region typography embedded directly in images without post-processing - **Subject-driven personalization** , HiDream-O1-Image consistently achieves strong performances across configurations with varying numbers of reference subjects. It handles up to 10+ reference objects in a single composition - **Instruction-based editing** , remove objects, swap backgrounds, change colors from a plain-text instruction - **Cinematic control** , natively supports 15 distinct shot types and camera perspectives for storyboard work The known weak spots: the original/full HiDream O1 model can show grid artifacts or other reference-image artifacts. In the upstream issue tracker, a HiDream developer recommends trying the Dev model because it should have fewer grid artifacts, and notes that reference-image generation is still being improved. The 1.5 closed-source variant is presumably more polished, but it's worth testing on your specific use case before committing. ## Where it sits in the competitive landscape HiDream.ai is a Chinese multimodal generative AI company building foundation models for image, video, 3D, and text generation. Products include HiDream-I1 (open-source text-to-image model, #1 on Artificial Analysis leaderboard), HiDream-E1 (image editing), vivago.ai (consumer AIGC platform), and PixMaker (marketing content). The company has raised 500M+ CNY (~$70M USD). The original HiDream-I1, released in early 2025, was a 17B sparse-MoE DiT operating in latent space , architecturally conventional, compete-on-quality. O1 is a reset: the parameter count goes down to 8B, the VAE and text encoder come out, and the architecture itself is the contribution. The 1.5 release continues that trajectory, now as a closed-source API product competing directly with GPT Image and Gemini. Parameter count is no longer a clean proxy for quality once architecture and data composition diverge. That's the assumption the field needs to update. HiDream's 8B open model already beats 56B FLUX.2 on five benchmarks. The 1.5 closed model now sits above Google's latest Gemini image offering in blind human preference votes. The VAE-free, unified-token-space approach is no longer a research curiosity , it's a production-grade architecture. ## How to use it HiDream-O1-Image-1.5 is a closed-source model priced at **$80 per 1,000 images**. It is currently available through HiDream's [HiHarness platform](https://hidream.ai) and the [Vivago platform](https://vivago.ai). If you want to experiment with the open-source base model first, HiDream-O1-Image is released under the MIT License and natively supported in ComfyUI (PR #13817). You can also run it via the [Hugging Face model page](https://huggingface.co/HiDream-ai/HiDream-O1-Image) or through fal.ai's API. For local inference, the FP8-quantized Dev checkpoint runs in roughly 10GB of VRAM, making it accessible on a single consumer GPU. The full model requires around 24GB. In general, the Full model is the better choice for realism and photographic detail. The Dev model is faster and often better for illustration, digital design, and cleaner grid/artifact behavior, but it can be more sensitive to scheduler and resolution choices. The practical use cases are broad: ad creative with embedded text, multilingual marketing graphics, storyboard generation for video pre-production, product visualization with consistent subject identity across shots, and any workflow where current models mangle the typography. The architecture's native multi-task design means you get editing and personalization from the same checkpoint , no pipeline switching required. ## Sources - [X post by @artificialanlys](https://x.com/artificialanlys/status/2064465242323869910) - [Artificial Analysis Text-to-Image Leaderboard](https://artificialanalysis.ai/image/leaderboard/text-to-image) - [HiDream-O1-Image Technical Report (arXiv:2605.11061)](https://arxiv.org/html/2605.11061v1) - [HiDream-ai/HiDream-O1-Image — Hugging Face](https://huggingface.co/HiDream-ai/HiDream-O1-Image) - [HiDream-O1-Image GitHub Repository](https://github.com/HiDream-ai/HiDream-O1-Image) - [HiDream-O1-Image-Dev: The 8B Pixel-Native Model That Beat 56B FLUX.2 — WaveSpeed Blog](https://wavespeed.ai/blog/posts/hidream-o1-image-dev-pixel-unified-transformer/)