What turbo tiers actually remove from a model
Fast, lite and mini variants cut sampling depth, resolution, reference slots or options, and each cut leaves its own artifact. Predict which shots survive.
A turbo tier is not the same model going faster. It is a different model, trained or configured to approximate the original, and something specific was given up to get there. The useful question is never "is turbo good enough" in the abstract. It is "which of the three things did this tier drop, and does my shot depend on that thing."
Four things get dropped, and you can usually tell which one from the catalog record before generating anything.
Where the tiers come from
Fast and turbo variants of familiar models generally come from distillation: training a smaller or faster model to imitate a larger one's outputs. That explains the behaviour. A distilled model is not making random mistakes, it is making the same kind of output as its parent with less capacity to resolve the hard parts, which is why turbo failures are consistent and predictable.
The catalog is the fastest way to see what a tier cost: compare variants in a family side by side and the deletions show up as narrower fields.
| Model | Credits | Resolution ceiling | Quality tiers | Durations | Frame rates | Reference slots |
|---|---|---|---|---|---|---|
| Seedance 2.0 | 69 | 1080p | 480p, 720p, 1080p, 4k | 4s–15s | not published | 9 image, 3 video, 3 audio |
| Seedance 2.0 Fast | 25 | 720p | 480p, 720p | 4s–15s | not published | 9 image, 3 video, 3 audio |
| Seedance 2.0 Mini | 16 | 720p | 480p, 720p | 4s–15s | not published | none published |
| Veo 3.1 | 40 | 4K | 720p, 1080p, 4K | 4s, 6s, 8s | 24, 30, 60 | not published |
| Veo 3.1 Fast | 15 | 4K | 720p, 1080p, 4K | 4s, 6s, 8s | 24, 30 | not published |
| Veo 3.1 Lite | 8 | 4K | 720p, 1080p, 4K | 4s, 6s, 8s | not published | not published |
Read the rows across and the pattern is clear. In the Seedance family, durations are untouched and the resolution ceiling halves at the first step down; the second step down keeps the resolution but drops reference handling entirely. In the Veo family, durations and the resolution ceiling are untouched all the way to Lite, and what erodes is frame-rate choice. Same idea, different sacrifice.
That is the whole method: find the field that got narrower, and you have found the cut.
The four cuts, and what each one leaves behind
Cut 1: sampling depth. The most common cut and the least visible in a spec table, because nobody publishes step counts. Fewer sampling steps means less refinement per frame.
Signature artifact: high-frequency detail goes soft and unstable. Fabric weave, hair strands, foliage, gravel, brushed metal, text on packaging. Freeze a frame and it looks acceptable; play it and the fine texture shimmers because each frame resolved it slightly differently. Fast motion and contact physics go next: feet slide, a hand passing in front of a face leaves a smear, water and cloth lose weight.
Shots that survive: anything with a dominant simple subject against a plain or shallow-focus background. Slow camera moves. Held expressions. Product on seamless.
Shots that do not: dense textured environments, running or dancing, anything where two surfaces make contact, and anything holding legible small text.
Cut 2: resolution ceiling. The most honest cut, because it is stated outright. Seedance 2.0 publishes quality options from 480p up to 4k against a 1080p output ceiling; both its Fast and Mini variants stop at 720p on both counts. You are not getting a soft 1080p, you are getting a hard limit.
Signature artifact: nothing at all on a phone preview, then a specific failure at real placement. Small text becomes unreadable, thin logo strokes break up, product detail that carried the shot disappears, and any crop compounds it because you are scaling a smaller source.
The trap is that upscaling afterwards is a different operation with a different failure mode. It adds plausible detail; it cannot recover a logo that was never resolved.
Shots that survive: anything vertical for feed, where the delivery target is modest and the subject is large in frame. Hooks, reaction cutaways, B-roll behind captions.
Shots that do not: anything with a product label, on-screen typography, a wide establishing shot with detail at distance, or a destination above feed resolution.
Cut 3: reference handling. The quietest cut and the one that surprises people, because it does not affect a single clip at all. It affects consistency across clips.
Seedance 2.0 and Seedance 2.0 Fast both publish the same reference configuration: up to 9 images, 3 videos and 3 audio references. Seedance 2.0 Mini publishes none. Those are very different tools. The first two can be handed a character sheet and a product angle set; the third takes a prompt.
Signature artifact: drift. The clip is fine, and the next clip is also fine, and the person in them is not quite the same person. Face structure shifts, a garment changes weave, a product's proportions move a few per cent, a colour that was locked in your brand kit reads warmer. On a single hero shot this is invisible. On a six-shot sequence it is the only thing anyone notices.
Shots that survive: one-off shots with no continuity obligation. Abstract texture, environment plates, anything where the subject is generic.
Shots that do not: any recurring character, any product hero, any sequence that cuts together, and anything where character consistency is the deliverable. If you have gone to the trouble of setting up reusable characters and products, a tier with no reference slots throws that work away.
Cut 4: option surface. Less a capability cut than a narrowing of what you can ask for, and the easiest to miss because everything still works.
Veo 3.1 publishes 24, 30 and 60 fps. Veo 3.1 Fast publishes 24 and 30. Veo 3.1 Lite publishes no frame-rate options at all. If you were planning to shoot at 60 for slow motion, two of those three tiers cannot do it and one of them will not tell you why. Aspect ratios narrow the same way in some families; Seedance 2.0 and its Fast tier both offer an adaptive option that the Mini tier drops.
Signature artifact: not an artifact, a rejection or a silent substitution. You ask for something the tier does not list, and either the call fails or you get the nearest thing it does support. The second outcome is worse, because it looks like success.
Predicting the outcome before you test
Run the shot through four questions.
- Does the frame contain fine repeating texture or hard contact physics? If yes, sampling depth is your risk. Test the turbo tier on that exact texture, not on a clean subject.
- Does anything in frame have to be legible or exact at delivery size? If yes, check the resolution ceiling against the placement, and remember that vertical crops shrink your usable pixels further.
- Does this shot have to match another shot? If yes, check reference slots. No slots means no consistency guarantee, regardless of how good the individual clip is.
- Am I relying on a non-default fps, aspect ratio or duration? If yes, check it is in the tier's published list rather than assuming it inherited from the parent.
Four "no" answers means the turbo tier is very likely fine and you should use it, because the credit difference across a family is substantial and the value shortlist exists for exactly this reason. One "yes" means test that specific property rather than the shot in general.
A note on sparse records
Some fast tiers publish less than their siblings. Hailuo 2.3 Fast carries no resolution ceiling, no quality tiers and no audio flag in the catalog, where Hailuo 2.3 Standard publishes 1080p and its quality options. That absence is not evidence of a cut, it is an absence of information, and the two are worth keeping apart. The right response is to test, not to assume. Kling 3 Turbo is the counter-example on the fields that decide this call: a 1080p ceiling, 720p and 1080p tiers, 3s to 15s in one-second steps, and audio declared off. Enough to place it without generating anything, even with its frame-rate list blank.
FAQ
Is the turbo tier of a new model better than the full tier of an old one?
Often, and that is the comparison worth running rather than turbo-versus-parent. A distilled version of a current-generation model frequently beats the previous generation's flagship on the axes that matter, at a fraction of the credits. Fast versus quality decision-making covers the framing; the point here is that generation gaps are usually larger than tier gaps.
Can I draft on turbo and final on the full tier?
Yes for composition, framing and pacing, which transfer cleanly. Not for anything the cut affects. If you drafted on a tier with no reference slots, the full-tier version with references will not match your draft. And a fixed seed does not carry meaning between two different models, so do not expect a turbo draft and a full-tier final to be the same shot with more detail. Plan for them to be siblings.
Do turbo tiers reroll more often?
In practice yes, on the shot types listed above, and that changes the arithmetic. A tier at a third of the credits that needs twice the attempts is still cheaper, but not by as much as the sticker suggests. Track your reroll rate per tier for a week before deciding a turbo tier is your default.
Why does the same variant name mean different things across families?
Because the naming is a marketing convention rather than a spec. "Fast" means a lower resolution ceiling in one family and fewer frame-rate options in another, and "lite", "mini", "turbo" and "flash" are not ordered consistently between vendors. Read the catalog record, not the suffix.