Models

Qwen-Image

Alibaba's other open image model, from the Qwen team rather than Tongyi-MAI. Heavier to run than Z-Image and more expensive per generation, with a mature editing and inpainting path.

Z-Image Turbo

Six billion parameters, runs in 16GB of VRAM, Apache 2.0. The cheapest credible image model to run in-house, and unusually good at legible text in a drawing. Geometry control is weaker than Flux.2.

Wan 2.7

Apache 2.0 open weights, which makes it the cleanest licence in video and the only serious option you can run on your own hardware. Strong on multi-shot narrative if you plan the sequence.

Flux.2 Dev

The open-weight tier of the Flux.2 family, sitting between Klein and Pro. A 32B model that wants serious hardware, in return for the strongest geometry retention you can run yourself.

Flux.2 Pro

The hosted tier. Better material realism than Klein, but you cannot train it on your own work.

Flux.2 Klein

Strong geometry retention and trainable on your own style. The current default for serious archviz.

Flux.1 Pro

The previous generation. Still widely used because the LoRA ecosystem around it is mature.

Flux3

Seedance 2.0

Strongest prompt adherence of the current video models and the best at holding a reference across shots. Reached mainly through ByteDance's own app rather than a broad developer API.

Seedream v4.5

Fast and cheap per generation. Good for volume option testing, less so for final imagery.

Nano Banana Pro

Reads a massing model unusually well and follows long prompts literally. Weaker on material realism.

Veo 3.1

The safest enterprise option and the strongest all-round realism, with audio generated alongside the frames. Every output carries a mandatory SynthID watermark.

Nano Banana

The standard tier. Faster and cheaper than Pro, with noticeably looser geometry.

Rodin Gen-2.5

The most detailed geometry of the hosted generators, with 4K textures and up to four reference views fused into one mesh. Priced accordingly.

Kling 3.0

The current generation of the Kling family, and the value pick among licensable models. Clips run to roughly fifteen seconds, with audio and stronger creative controls than 2.5.

Kling 2.5

Previous generation. Cheaper per second, with more drift on facade detail through camera moves.

Kling

Turns a still render into a credible camera move. Watch for drift on fine facade detail.

LTX-2.3

The longest single-pass clip of the current models at around twenty seconds. Apache 2.0, trained on licensed data, and free below a revenue threshold, which is a rare combination worth reading closely.

Ray 3.2

Luma's current model, behind the Dream Machine interface. Quick and cinematic on short image-to-video clips, with a storyboard mode for stitching a longer sequence.

Meshy 6

The most complete hosted pipeline: triangle or quad topology, target polycount, symmetry and remesh controls. Commercial use sits behind the paid tier.

TRELLIS.2

Single reference image in, mesh with native PBR out. An MIT licence and self-hosting make it the least encumbered option for work you intend to sell.

Midjourney

Unmatched for atmosphere and mood boards; poor at holding your actual geometry.

Hailuo 2.3

The fastest turnaround and the lowest price of the hosted models, which makes it the sensible choice when you expect to discard most attempts. Quality sits below the leaders.

GPT Image 2

The first OpenAI image model that reasons through a prompt before drawing. Unmatched for legible text inside an image — signage, drawing titles, plan callouts — and takes up to sixteen reference images for in-context editing.

GPT Image 1.5

The December 2025 tier, still in the API and faster than 2. Good text rendering, but without the reasoning pass, so complex layout instructions have to be spelled out.

Sora 2

Strong realism but no character or reference control. OpenAI has announced it is winding Sora down — the web and app experiences ended in April 2026 and the API is scheduled to follow in September. Do not build a workflow on it.

Runway Gen-4.5

Reference-image controls and a storyboard editor make it the most directable of the hosted models, which matters more than raw score for client deliverables. Topped the no-audio leaderboard at launch and has since been overtaken.

Stable Fast 3D

Sub-second meshes, which makes it the only one of these usable in a live loop. Albedo only, no PBR, and the geometry is a rough block rather than a finished asset.

Hunyuan3D 2.1

Turns a photograph or elevation into a textured mesh, with real PBR material maps rather than baked colour. Weights and training code are both public, so it can run inside the office.

HunyuanImage 3.0

Tencent's open image model, competitive with the Flux and Qwen open tiers. Heavy hardware requirements compared to Z-Image for a similar result.

HunyuanVideo 1.5

Tencent's open video model, Apache 2.0. Behind the hosted leaders on quality, but the licence and self-hosting make it viable where client footage cannot leave the office.

Tripo 3.0

High polygon counts and quad-mesh output, with multi-image input when one view is not enough. Slower per generation than its own 2.5, which remains the faster iteration tier.