← Back to feed

Every major text-to-image model in production — FLUX, SDXL, Stable Diffusion 3, Imagen — runs the same basic stack: a frozen text encoder routes language into a latent diffusion transformer, a VAE compresses images into compact latent tokens, and a decoder converts them back. HiDream threw out the latent stack entirely.

HiDream-O1-Image (codename Peanut) runs on a Pixel-level Unified Transformer (UiT) that processes raw pixel patches, text tokens, and task-condition tokens in a single shared sequence. No VAE. No disjoint text encoder. The diffusion process happens in pixel space from start to finish.

What Shipped

The release on May 8, 2026 includes two checkpoints and an inference agent:

• HiDream-O1-Image (full): 8B parameters, 50 inference steps, CFG guidance scale 5.0, up to 2,048 × 2,048 output
• HiDream-O1-Image-Dev: Same 8B, 28 steps, CFG guidance scale 0.0 (guidance-distilled), meaningfully faster with a quality trade-off
• Reasoning-Driven Prompt Agent: A standalone script (prompt_agent.py) that takes a raw user instruction, reasons through layout, subject attributes, physical logic, and text-rendering requirements using Gemma-4-31B-IT as backend, then outputs a structured JSON with a refined prompt, reasoning trace, and resolved implicit knowledge. Feed the rewritten prompt into the inference pipeline.

Both model variants are MIT licensed. Commercial use permitted with no restrictions. The Dev weights and a hosted Hugging Face Space are live.

Architecture Rationale

Pixel-space generation has historically been more expensive than latent diffusion. Pixels carry more information than compressed latent tokens, so compute per token rises. HiDream’s answer is sparsity and scheduling — the Dev variant gets guidance distillation to skip the doubled compute that classifier-free guidance normally requires. At 28 steps without CFG, Dev is faster per image than many 50-step latent models.

The payoff comes in categories where latent models struggle: long-text rendering, complex multi-element composition, and multilingual text in image. Latent VAEs compress away fine detail that matters for accurate characters and precise spatial layout. Running in pixel space preserves it.

HiDream benchmarks O1-Image across five suites: GenAI-Bench (compositional generation), T2I-CompBench++ (dense prompt alignment), ELO-style human preference, long-text rendering, and multilingual text rendering. On compositional tasks and text rendering specifically, the model reports parity with or superiority over larger open DiTs and select closed-source commercial models.

These numbers have not been independently replicated at scale. The Artificial Analysis Text-to-Image Arena ranking (#8 overall, highest open-weight) is an independent measure based on human preference voting — it says the model produces images people prefer, not just images that score well on automated suites.

Open-Weight Context

At #8 on the AA Text-to-Image Arena, O1-Image enters above every prior open-source model at launch. Closed-source competition at the top of the board includes GPT Image 2, Imagen 4, and FLUX Pro variants that run as API-only. O1-Image is the first open-weight entry to reach the top-10 tier on AA at release since FLUX 1.1 Pro Ultra.

The practical ceiling: local inference requires roughly 35GB of VRAM for the full model. Consumer 3090/4090 hardware is out. The Dev variant quantized may be more tractable, though official quantized weights haven’t shipped yet. Cloud inference is the realistic path for most production workloads until community quantization arrives.

What Matters

HiDream’s prior release — HiDream-I1, a 17B sparse-MoE DiT in latent space — attracted significant developer adoption in early 2025. O1-Image is a harder architectural bet: fewer parameters, removed components, higher conceptual clarity. The naming convention is deliberate — O1 signals the integrated reasoning agent, even though the diffusion sampler itself is a standard one-shot pipeline.

The Reasoning-Driven Prompt Agent is the immediately practical piece. Resolving ambiguous user prompts before generation — understanding that “a Tang Dynasty general” implies specific armor and weapons, for instance — addresses one of the most consistent failure modes in text-to-image production. The agent uses Gemma-4-31B-IT, which carries its own HuggingFace licence, adding a wrinkle for teams that want the full pipeline under MIT.

If the pixel-space approach holds up under independent testing and the community ships quantized variants, O1-Image represents a structural step toward simpler image pipelines — one transformer, one training paradigm, one licence.

Key Numbers

• Parameters: 8B (both full and Dev checkpoints)
• Architecture: Pixel-level Unified Transformer (UiT), no VAE, no disjoint text encoder
• Output resolution: up to 2,048 × 2,048
• Inference steps: 50 (full), 28 (Dev)
• AA Text-to-Image Arena rank: #8 overall, #1 open-weight
• License: MIT (commercial use permitted)
• Release date: May 8, 2026
• Prompt agent backend: Gemma-4-31B-IT (requires separate Gemma licence)