What Turbo changes
Turbo keeps the Qwen-Image-2.1 model and prompt format but ships with its own 8-step sampling schedule, which the pipeline loads automatically. Fewer steps mean faster images. Guidance is fixed at CFG 1, and prefix KV caching reuses the text and reference-image context across steps.
The fast box or standard Qwen Image 2.1?
- Fast box (2 credits): quick text-to-image drafts; a 1312 × 1312 image took about 11 seconds end to end in our test.
- Standard Qwen Image 2.1 (1 credit): generation and reference-image editing in the main studio.
- Prompts carry over unchanged, so you can draft in one and finish in the other.
Running it locally
The checkpoint loads with QwenImage21Pipeline in Diffusers. It needs the latest Diffusers source (support for pipeline-configured sampling sigmas arrived in PR #14950), a recent Transformers release, Accelerate and a CUDA build of PyTorch. Setting num_inference_steps alone does not change the saved schedule; only an explicit sigmas argument overrides it, and other schedules have not been evaluated. Follow the model card for the exact install and loading code.
Checkpoint resolution presets
- 1:1 — 2048 × 2048
- 4:3 — 2400 × 1792 and 3:4 — 1792 × 2400
- 3:2 — 2528 × 1696 and 2:3 — 1696 × 2528
- 16:9 — 2752 × 1536 and 9:16 — 1536 × 2752
Using the fast box on this site
The box at the top of this page sends your prompt, unchanged, to fal’s hosted Qwen Image 2.1 API (fal-ai/qwen-image-2.1). fal does not publish which Qwen Image 2.1 checkpoint or how many steps it runs, so treat it as a fast hosted Qwen Image 2.1 option rather than a guaranteed copy of the 8-step Turbo checkpoint.
Each image costs 2 credits, is about 1.5–1.8 megapixels in your chosen ratio (1312 × 1312 at 1:1, 1664 × 928 at 16:9, 1536 × 1152 at 4:3) and is saved to your history. If the provider refuses a request or it fails, the credits are returned automatically. To edit, add one or two reference images in the box: 3 credits with one reference, 4 credits with two. References are scaled down to at most 1 megapixel before they are sent, because fal also bills reference pixels. For three references, use Qwen Image 2.1 Edit in the main studio (1 credit per edit).
For reference, fal lists this API at $0.02 per megapixel for generation and $0.0367 per megapixel for edits, rounded up to whole megapixels (third-party API list price, checked on 2026-10-27). That is the provider’s price, not ours.