MODEL PROFILE · HiDream.ai

HiDream-O1-Image

A source-linked guide to the original full O1 Image model, its hosted controls and megapixel billing. The newer 1.5 homepage label is kept separate.

Source reviewed: 2026-10-02 · Image model · open weights

Provider documentation summary. No hands-on results or independent ranking are claimed.

Documented facts

The following fields come from the official provider source, reviewed on the date above.

Exact checkpoint: HiDream-ai/HiDream-O1-Image, the full undistilled model. Dev and Dev-2604 are separate checkpoints.
Dated release record: The repository records the original 8B full and Dev open-weight release on May 8, 2026. This October catalog addition is not a new launch.
Documented tasks: Text-to-image, instruction-based editing and multi-reference subject personalization, with output up to 2048×2048.
Implementation and controls: The repository documents 50 steps for the full model and recommends it for editing. A CUDA-capable GPU is required by the local instructions.
Claimed strengths: The developer emphasizes visual text, layout and subject preservation. These are proposed evaluation areas, not verified quality advantages.

Open weights and license scope

The official model card labels this checkpoint MIT and states that the model and repository code use that license. Its optional local prompt-refinement backend has a separate Gemma license. Record the chosen refiner as well as the image checkpoint when reproducing a result. HiDream model card and licensing notes.

Hosted API: inspect the actual schema

fal documents fal-ai/hidream-o1-image with a required prompt: no reference images for generation, one for editing, multiple for personalization. Requested dimensions round down to multiples of 32, up to about 2048×2048. Defaults include 50 steps and guidance 5; a seed, image count and PNG/JPEG/WebP export are exposed. The single-reference mode can preserve the input aspect ratio. These are documented controls, not results from a call. No maximum reference count was verified. fal endpoint input schema.

Pricing uses megapixels, not a flat image fee

The fal page displays $0.01 per megapixel as of October 2. Assuming one megapixel means 1,000,000 output pixels with no minimum, rounding or other charges, a 1024×1024 image calculates to about $0.01049 and a 2048×2048 image to about $0.04194. These are arithmetic estimates, not observed invoices; confirm billing rules before budgeting. The displayed rate does not establish self-hosting cost. fal displayed endpoint price.

Why this is not a 1.5 or 2.0 specification sheet

The homepage headline advertises O1 Image 1.5, but its chart note explicitly attributes the plotted results to the open-source 8B O1 Image. This review does not establish that the repository checkpoint or fal route is version 1.5. The reviewed sources also do not establish an O1 Image 2.0 release. Those version identities remain separate verification tasks. HiDream homepage and chart attribution.

The practical decision

Choose the task before choosing the access route. For a packaging edit, define which lettering and surfaces must remain intact. For subject personalization, decide whether likeness, clothing or product geometry is the acceptance criterion. For a new illustration, list the required objects and their relationships. Mixing these tasks into one overall preference vote hides the failure that matters to the project. A hosted route can simplify execution, while a local checkpoint gives a concrete artifact to record; neither choice by itself proves output quality.

What to record in your own comparison

Save the original instruction and any refined prompt separately. If one candidate receives a rewritten prompt and another receives the raw instruction, record that difference instead of attributing all improvement to the image model.
Compare at the same delivered dimensions and inspect small text at readable scale. For edits, mark the requested change and inspect the untouched regions too; a pleasing replacement can still discard required details.
Log checkpoint or route, reference order, settings, output size and accepted versus rejected images. Use the total cost of attempts per accepted result for your own budget, rather than assuming every returned image is usable.

Limitations and unknowns

No inference calls, paid generation or new samples were produced for this profile. We did not measure speed, fidelity or hardware memory requirements. Provider examples and published evaluations are not DualView tests. Hosted deployment revision, exact billing rounding, prompt-length ceilings and a maximum reference count remain unverified. Do not copy these specifications to a differently named version.

Compare your saved results

Use the DualView editor to inspect files you already have. The accepted-output calculator uses your own rates and counts. Neither link runs a model or purchases generation.

Related reporting

Related model profiles

All model profiles and dated changes · Subscribe via RSS