Lightricks shipped HiDream-O1-Image-Dev-2604 on April 26, 2026. The model is an open-weight text-to-image system from the team behind LTX Video and popular creator apps. It lands on Hugging Face with an Apache 2.0 license for weights and no per-image API fees. The release targets local inference, fine tuning, and commercial use without vendor lock-in. The model uses a diffusion transformer with a paired text encoder and VAE. It supports a 512 token prompt context, up from 77 tokens in older CLIP based models. That means longer prompts can hold more scene detail. It is a direct answer to closed image APIs that charge per generation. For a wider look at free image tools, see this comparison of free AI image generators.

Lightricks built the model for practical creative work. Total parameters sit at 8 billion across the transformer, text encoder, and VAE. The Dev variant is the open release. A larger Pro version may stay closed, but the Dev weights ship with no access token. The benchmark profile is striking. HiDream-O1-Image-Dev-2604 scores 0.73 on GenEval and 87.6 on DPGBench, numbers that place it above several paid APIs in prompt fidelity. It also wins 62 percent of blind human preference tests against Stable Diffusion 3.5 Large on a 1,000 image set. Those are not absolute wins across every style, but they matter for open models. You can read more about open versus paid model economics in this AI image pricing analysis.

Why it matters is simple. Most high quality text-to-image systems are locked behind paid APIs or restrictive non-commercial licenses. HiDream-O1 flips that. The Apache 2.0 license covers weights, so you can self-host, fine tune, distill, or use outputs commercially without paying a per-image tax. That changes budget math for startups and independent creators. A single A100 or RTX 4090 can serve dozens of users when quantized to 4 bit. The model does not need a cloud round trip, which also keeps prompts private. For developers watching the free tier squeeze across AI tools, this release is a rare counter trend. See the broader AI free tier landscape for context.

The official announcement points to Lightricks as the source. Files are hosted on Hugging Face for direct download, with safetensors weights and a reference inference script. You do not need a special repo link to start. Search for the model name on the hub or visit the Lightricks site. The release date is April 26, 2026. This follows a wave of open image and video launches, but few combine this license, context length, and local footprint. It also arrives as Google and OpenAI adjust image pricing and free tier access. If you want a broader list of no-cost models, check the best free AI models in 2026.

How Do the Top Options Compare?

Model License Parameters Prompt Limit Strongest Use Case
HiDream-O1-Image-Dev-2604 Apache 2.0 8B total 512 tokens Commercial local generation and fine tuning
FLUX.1-dev Non-commercial 12B 512 tokens Aesthetics and high detail
Stable Diffusion 3.5 Large Stability AI Community 8B 77 tokens Typography and open community tools
Google Gemini 2.0 Flash Image Proprietary API Not disclosed Context dependent Fast cloud generation with free tier limits

Specs reflect fp16 or vendor listed values. Cloud prompt limits may shift with API updates; check vendor pages before building.

1. HiDream-O1-Image-Dev-2604 , Open-source commercial image generation without per-image fees

Lightricks released HiDream-O1-Image-Dev-2604 as an Apache 2.0 open-weight model. The package includes a diffusion transformer, a text encoder, and a VAE. Total size is about 8 billion parameters. Weights come in fp16 at roughly 16 GB. A 4 bit quantized version drops to about 8 GB, making it usable on a 12 GB RTX 3060 or 4070 with offloading. The model supports a 512 token text prompt, which is a large jump from the 77 token ceiling in many Stable Diffusion models. That longer context lets you describe multiple subjects, lighting, and composition without relying on ComfyUI prompt tricks. You can find the official source through Lightricks and the files on Hugging Face.

Benchmark numbers from the release are strong. It records 0.73 on GenEval and 87.6 on DPGBench. Those scores place it above several paid image APIs on text rendering and attribute binding. In a 1,000 image blind test, human raters preferred HiDream-O1 over Stable Diffusion 3.5 Large 62 percent of the time. These are not perfect scores. Complex hands, long text strings, and rare camera angles still fail. But the gap between open and closed models has narrowed enough for production use. For creators tired of API credits, this is a meaningful shift. Compare it with other free options in the free AI image generator guide.

The license is the main event. Apache 2.0 allows commercial use, modification, redistribution, and fine tuning. You can build a paid product on top of the weights without revenue sharing. That is rare for a model with this benchmark profile. The release also includes a simple Gradio demo script. You do not get a hosted API from Lightricks with the Dev release, so you must rent or own a GPU. That is a downside for non technical users. For developers, the total cost of serving can be lower than API volume pricing after a few hundred images per day. See how usage based pricing has hit AI coding tools for a parallel lesson in how free tiers change.

Key strengths:

  • ✅ Apache 2.0 license removes commercial and fine tuning restrictions
  • ✅ 512 token prompt context captures longer scene descriptions
  • ✅ Quantized weights run on consumer 12 GB GPUs
  • ✅ Benchmarks beat several paid APIs on GenEval and DPGBench
  • ✅ No per-image fee after self-hosting costs
  • ❌ No hosted API from Lightricks for the Dev variant
  • ❌ Quantized mode can lose small text fidelity
  • ❌ Requires technical setup for ComfyUI or Diffusers inference

Who it’s for: Developers and creators who want a commercially safe, self-hosted text-to-image model with modern prompt limits.

2. FLUX.1-dev , High detail aesthetics when commercial use is not required

FLUX.1-dev from Black Forest Labs set a high bar for open image quality in 2024. It uses a 12 billion parameter rectified flow transformer and a T5 text encoder. The prompt context reaches 512 tokens. Visual detail, skin texture, and complex lighting are its strong points. Many community fine tunes, LoRAs, and ComfyUI workflows target FLUX. The model is available on Hugging Face under a non-commercial license. That license restricts revenue generating use for the dev variant. A separate Schnell model is Apache licensed but trades some quality. For a cost view of image APIs, see this AI image pricing breakdown.

FLUX dev is not free for commercial work. If you run a business, you need a separate license or switch to Schnell. That is the main reason HiDream feels disruptive. Many teams adopted FLUX for prototypes, then paid for API access when shipping. The local footprint is also heavy. FP16 weights need about 24 GB of VRAM unless you quantize to 4 bit. A 16 GB card can run quantized versions, but output speed drops. Community optimizations help. If you want a free model for paid projects, HiDream’s Apache license is simpler. See where open models fit in the best free AI models of 2026.

Still, FLUX has maturity. The ecosystem of ControlNets, IPAdapters, and style LoRAs is deeper than a brand new model. That matters if you need precise pose control or consistent character pipelines. HiDream will take time to build that tooling. FLUX also handles long text prompts well, but occasional garbled letters remain. For pure visual polish on non-commercial projects, FLUX dev often wins aesthetic comparisons. For any revenue generating product, read the license carefully before using dev weights.

Key strengths:

  • ✅ Excellent visual detail and lighting control
  • ✅ Broad ecosystem of LoRAs and ControlNets
  • ✅ 512 token prompt context
  • ❌ Non-commercial dev license blocks free business use
  • ❌ FP16 weights need high end GPUs without quantization
  • ❌ Slower than smaller models on midrange hardware

Who it’s for: Non-commercial tinkerers and researchers who value maximum image fidelity and already know ComfyUI.

3. Stable Diffusion 3.5 Large , Typography and open community workflows

Stable Diffusion 3.5 Large from Stability AI is an 8 billion parameter model under the Stability AI Community License. It improves text rendering and follows multi subject prompts better than SDXL. The license is free for non-commercial use and for commercial use under one million annual revenue. Above that, Stability expects a paid license. Prompt context is about 77 tokens through the CLIP text encoder, though some integrations use a second encoder. This model is everywhere in ComfyUI, Automatic1111, and Forge. The open tooling is a real advantage. For creators watching provider pricing changes, the model remains a stalwart. Check the AI free tier landscape for updates.

Compared with HiDream, SD3.5 Large has a weaker context window. Long prompts get truncated, which forces users to abbreviate or use regional prompting. Text rendering has improved but still lags newer models on small logos and quote marks. Human preference rates it lower than HiDream in the April 2026 blind test. However, the community knowledge base is extensive. You can find thousands of tutorials, embeddings, and fine tunes. That can outweigh benchmark gaps for fast iteration. If you run a small business under the revenue threshold, SD3.5 Large is a safe choice. The license is not fully open by OSI standards. For a detailed guide on free image tools, see the free AI image generator comparison.

Hardware requirements are moderate. FP16 weights need around 16 GB of VRAM. Quantized GGUF versions run in 8 GB. The model is stable and well integrated with diffusers. Fewer breaking changes hit SD3.5 workflows compared with bleeding edge releases. That reliability matters for production. Still, the 77 token limit is a real ceiling for detailed scene prompts. If you regularly write long prompts, HiDream’s 512 token context removes that friction.

Key strengths:

  • ✅ Strong typography and logo generation
  • ✅ Huge ecosystem of fine tunes and tools
  • ✅ Community license allows small business use
  • ❌ 77 token context truncates long prompts
  • ❌ Not fully open source by OSI definition
  • ❌ Benchmark text fidelity still lags newer models

Who it’s for: Creators who need mature ComfyUI workflows and can accept the one million dollar revenue cap.

4. Google Gemini 2.0 Flash Image , Fast cloud image generation with free tier quota

Google Gemini 2.0 Flash Image is a closed API model. It is not open source and has no public weights. The free tier includes a limited number of images per day or month, though exact limits shifted in 2026. Gemini Flash Image is fast and handles conversational edits well. You can ask it to modify a previous image in chat. That is useful for rapid prototyping. Google has also cut prices on some AI plans, but image generation still consumes quota. For a full analysis, read the Google versus OpenAI image pricing breakdown. The official source is Google AI.

Closed models like Gemini Flash Image compete on convenience. There is no GPU setup, no ComfyUI, no Python environment. You send a prompt and receive a file. For non technical users, that is a genuine advantage. The downside is lock-in. You cannot inspect the weights, fine tune the model, or guarantee prompt privacy. Free tier changes can remove access without notice. Several providers tightened free image quotas in June 2026. If the free tier disappears, your workflow stops. That is why open models matter even when they require setup. See the major provider free tier adjustments for the latest.

Gemini Flash Image has strong instruction following and a wide style range. It handles text rendering well in Google’s own demos. But independent benchmarks show open models are closing the gap. The April 2026 HiDream release equals or beats it on GenEval for some prompt categories. That is notable because Gemini is a paid API at scale. If your volume is low, the free tier is attractive. If your volume grows, the per-image price dominates. Self-hosting an Apache licensed model shifts the cost curve. For developers, the choice is not purely technical. It is about who controls the weights. Read more about open model economics.

Key strengths:

  • ✅ No local GPU setup required
  • ✅ Fast generation and conversational editing
  • ✅ Free tier works for low volume use
  • ❌ Closed weights prevent fine tuning and self-hosting
  • ❌ Free tier limits can change without notice
  • ❌ Per-image pricing becomes costly at scale

Who it’s for: Users who want instant cloud images without technical setup and can accept API lock-in.

Frequently Asked Questions

Is HiDream-O1-Image-Dev-2604 free for commercial use?

Yes. The weights use an Apache 2.0 license. You can use outputs in paid products, fine tune the model, and redistribute your modifications without paying Lightricks. The license does not cover the Lightricks trademark or any hosted service they may sell, but the weights themselves are free for commercial use.

What GPU do I need to run HiDream-O1-Image-Dev-2604?

FP16 weights need about 16 GB of VRAM. An 8 bit quantized version fits in about 10 GB, and 4 bit can run on a 12 GB card with some offloading. Expect slower generation on lower end hardware. A 24 GB card gives the best balance of speed and quality.

How does HiDream-O1 compare to closed models like Gemini Flash Image?

On GenEval and DPGBench, HiDream scores 0.73 and 87.6 respectively. This places it above several paid APIs in prompt fidelity. Closed models still win on convenience and sometimes on long text rendering. But the gap is small enough for local production use.

Where can I download the weights?

Download safetensors weights from Hugging Face. Search for the model name or follow the link from the Lightricks site. You do not need an access token for the Dev weights. A reference Gradio script is included.

Does it support LoRA fine tuning?

Yes. The Apache license permits fine tuning and LoRA training. The model works with common diffusers based trainers. Memory requirements depend on rank and batch size, but a 16 GB GPU can train small LoRAs with gradient checkpointing.

What makes the 512 token prompt limit useful?

Most open image models truncate prompts to 77 tokens. A 512 token context lets you describe multiple subjects, backgrounds, lighting, and composition in one prompt. It reduces the need for regional prompting tricks and improves adherence to complex scenes.

What Should You Remember?

  • Open weights: HiDream-O1-Image-Dev-2604 ships under Apache 2.0, so commercial use costs zero per image.
  • 512 tokens: The prompt context is six times longer than Stable Diffusion 3.5, so complex scenes survive.
  • Local first: Quantized weights run on a 12 GB consumer GPU and avoid API privacy risks.
  • Benchmarks: GenEval 0.73 and DPGBench 87.6 beat several paid image APIs.
  • License matters: Unlike FLUX.1-dev non-commercial terms, HiDream has no revenue cap.
  • Tradeoff: No hosted API means you must run your own GPU or use a community service.
  • Ecosystem gap: New model tooling is thinner than FLUX or SD3.5, so factor in setup time.

Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.