When to use a fast model, when to pay for the full one, and how to set each up

Find out what kind of fast it is

"Fast" in a name can mean three different things. A distilled model is a separate student trained to finish in 1 to 8 steps: SDXL Turbo, SDXL-Lightning, FLUX.1 schnell, FLUX.2 klein and Z-Image Turbo all are, and their model cards say so.

A lighter tier is a smaller or cheaper version whose maker doesn't say how it is made: Google's Nano Banana 2 Lite and Veo 3.1 Fast, Seedance 2.0 Fast, Kling 4.0 Flash. Same model, faster server: Midjourney's Turbo mode runs the same model on faster GPUs and charges double Fast time. Each behaves differently, so check before you rely on the name.

I'm looking at a model called [MODEL NAME] in [APP]. Tell me which kind of "fast" it is: 1) a distilled few-step model (trained to copy a slower teacher in a few steps), 2) a lighter or cheaper tier whose method isn't published, or 3) the same model run on faster hardware or at lower resolution. Only say it's distilled if the maker's model card or docs say so; otherwise say it's unknown. Then tell me what that means for quality, variety and settings.

Use fast for exploring

Fast models earn their place when you need many tries: testing a prompt, trying twenty compositions, storyboards and thumbnails, mood frames for a client call, or anything live and interactive. At a few steps you see results almost as quickly as you can type.

Most of these models drop classifier-free guidance, the setting that sharpens how closely the image follows the prompt, and with it negative prompts. That is fine for exploring: you are looking for a direction, not a final.

Switch to full for the final

Go to the full model for the image people will actually see: the hero shot, close-up faces and hands, small text, product detail, and any batch where you need the images to differ from one another. Variety is the most consistent loss. Tongyi rates Z-Image Turbo's diversity "Low" against "Medium" for the full Z-Image.

Also use the full model when you need negative prompts, strong guidance or fine-tuning. Stability AI says SDXL Turbo cannot render legible text and is fixed at 512 × 512, so it is a sketching tool, not a finishing one.

Set the fast model up correctly

The commonest mistake is running a distilled model with normal settings: 30 steps and guidance 7 is slower and tends to overcook the image. Use the maker's numbers. SDXL Turbo: 1 to 4 steps, guidance 0. SDXL-Lightning: the checkpoint that matches your step count (1, 2, 4 or 8), guidance 0, and in ComfyUI the Euler sampler with the sgm_uniform scheduler.

FLUX.1 schnell: 1 to 4 steps, guidance 0. FLUX.2 klein 4B: 4 steps, guidance 1.0. Z-Image Turbo: 8 model passes (9 steps in the Diffusers example), guidance 0. LCM-LoRA on SDXL: 2 to 8 steps, guidance 0 or between 1 and 2, with the LCM scheduler. The full list is under "Put it to work".

I'm running [MODEL] in [ComfyUI / Forge / Diffusers / fal / Replicate]. Give me the settings the model card recommends: number of steps, guidance or CFG, sampler and scheduler, and resolution. Tell me which of my current settings to change: [PASTE YOUR SETTINGS]. If you don't know the model card's values, say so instead of guessing.

Explore fast, then finish full

Work in two stages. Stage one: run your idea through the fast model in batches, change one thing at a time and keep the best two or three. Stage two: take the winning prompt, unchanged, to the full model, and attach the fast draft as an image reference if the full model accepts one.

Don't expect the same seed to give the same picture on the other model. A seed only repeats an image on the same model, at the same size and settings. The prompt and the reference image are what carry the idea across.

Here is the fast draft I picked and the prompt that made it: [PROMPT]. I'm going to make the final in [FULL MODEL]. Rewrite the prompt for that model: keep everything that made the draft work, add the detail the fast model left vague (materials, light direction, lens, small text exactly as it should read), and tell me whether to attach the draft as an image reference. Keep it under 80 words.

Test both on your own prompt

Claims about quality are averages. Run four images on the fast model and four on the full one with the same prompt, then compare. You will quickly learn where the fast model is good enough for your kind of work, and where it falls short.

Look for three things: how different the four images are from each other, fine detail at 100% (skin, fabric, hair, text), and whether every part of the prompt made it in.

Here are four images from [FAST MODEL] and four from [FULL MODEL], same prompt: [PROMPT]. Compare the two sets on: variety between the four images, fine detail at full size (skin, fabric, hair, text), and prompt adherence (list any part of the prompt missing in each set). Then tell me for which of my jobs the fast model is good enough: [LIST YOUR JOBS, e.g. storyboards, Instagram posts, client finals].

2 Lite · 2.8 s

2.1 · 10.8 s

Pro · 12.4 s

Do the same for video, where it saves more

Video tiers follow the same logic and the savings are larger, because every second costs. Use the fast or draft tier to test motion, timing and the camera move, then render the keeper on the full tier.

As of October 2026: Veo 3.1 Fast makes the same 4, 6 or 8 second clips as Veo 3.1 but tops out at 1080p, where Veo 3.1 goes to 4K. Runway added a Seedance 2.5 Draft mode that makes 480p drafts. Partners such as fal list Kling 4.0 Flash as the speed version of Kling 4.0.

Few-step open models: steps and guidance (from each model card)

SDXL Turbo (Stability AI): Adversarial Diffusion Distillation, 1 to 4 steps (one is enough), guidance 0, 512 × 512, no legible text. SDXL-Lightning (ByteDance): progressive adversarial distillation, checkpoints for 1, 2, 4 and 8 steps (2, 4 and 8 are the quality ones), guidance 0, trailing timestep spacing, Euler with sgm_uniform in ComfyUI.

FLUX.1 [schnell] (Black Forest Labs): latent adversarial diffusion distillation, 1 to 4 steps, guidance 0, Apache 2.0. FLUX.2 [klein] 4B: "our fastest distilled model", 4 steps, guidance 1.0, Apache 2.0, about 13 GB of VRAM. Z-Image Turbo (Tongyi-MAI): Decoupled-DMD distillation, 8 model passes, guidance 0, Apache 2.0. LCM-LoRA for SDXL: 2 to 8 steps, guidance 0 or 1 to 2, LCM scheduler.

For comparison, the full models: FLUX.1 [dev]'s example uses 50 steps at guidance 3.5, and the full Z-Image uses 50 steps with guidance. FLUX.1 [dev] is itself "guidance distilled" (no separate unguided pass), but not step-distilled.

Fast and full pairs in the apps (as of October 2026)

Google: Nano Banana 2 Lite ("the fastest and cheapest Gemini image model"), Nano Banana 2.1 ("Flash-level speed and cost efficiency") and Nano Banana Pro ("the premium choice for the most complex visual tasks"). Google does not say how its Flash models are made, so treat them as a lighter tier, not as a known distilled model.

Midjourney: Draft mode is for quick iteration. In V8.1 and V8.2 one Draft prompt makes 24 images at 512 × 512 for 0.4 GPU minutes. Turbo mode is the same model on faster GPUs, up to four times faster than Fast at double the GPU time, and isn't supported in V8.1. Ideogram: render speeds Flash (where supported), Turbo ("rougher or more sketch-like"), Default and Quality.