Choose a video model
Pick the model that renders a shot. Ask first: does the person know what the shot needs, and are the shots composed yet? If not, draft on H3 Max (below). If they do, match the shot's hardest demand (resolution, realism, motion, an audio or end-frame input) to the tables.
Prices and timings change. The tiers below are as of 2026-10; get exact prices from
describe(model) (or run with estimate) before a batch, and call describe before
the first run of a model in a job.
The default: draft on H3 Max, then upgrade
When the person doesn't know, or the shots aren't composed yet, render every shot on H3 Max first. It is fast and cheap, with no minutes-long wait per revision, so composition and timing can be iterated with the person. Then upgrade only the shots whose complexity demands it: 1080p, photographic realism, close-up texture, or complex physical motion go to Seedance 2.5.
A draft approves composition and timing, not motion. Review the motion again on the upgraded render; a different model moves differently.
The models
H3 Max (MiniMax)
fal/minimax/h3-max/image-to-video, …/reference-to-video, …/recast
| Time per render | The fastest tier: quick enough to iterate in conversation |
| Top resolution | 1080P (768P plus latent refinement); 768P and 480P cost less. Priced per second by tier, whatever the canvas shape |
| Look | Plastic: facial hair, skin texture and hair strands come out mottled and smoothed, painted rather than photographic. Viable wide, or when the subject has little fine texture |
| Motion | Good motion and performance; strong, cinematic blocking. Not complex physical motion |
| Inputs | First frame; end frame (i2v and r2v); up to 9 references (r2v); audio: target_audio_url lip-syncs a still (i2v), a reference track up to 15 s (r2v). The only model that takes an end frame and audio together. Lengths requested in seconds, from about 1 s up to 15 s |
| Cost tier | Low (768P is half the price of 1080P) |
| Leaves | h3-max-image-to-video, h3-max-dialogue, h3-max-lipsync-a-still, h3-max-recast |
Renders come back a little longer than asked (ask for 1 s, get 1.2–1.4 s), so a short shot renders at its own length and is trimmed by a few frames. In r2v it pushes in by default, can cut to new angles on long lines, and can duplicate a person.
Seedance 2.5 (ByteDance)
fal/bytedance/seedance-2.5/us/image-to-video, …/reference-to-video
| Time per render | Slow: 720p ≈ 6 min, 1080p 6–30 min |
| Top resolution | 1080p native (720p plus Topaz only to save money; it softens detail) |
| Look | Photographic, with real grain; holds close-up texture |
| Motion | Complex physical motion (papers bursting and falling like feathers); literal, restrained blocking; holds a locked frame |
| Inputs | i2v: first and end frame, no references, no seed, no aspect ratio (taken from the image). r2v: up to 29 reference images and an audio track up to 30 s, but no frame inputs (the still is @Image1 by the prompt) |
| Cost tier | High: the bulk of a job's render spend |
| Leaves | seedance-2-5-image-to-video, seedance-2-5-dialogue |
i2v opens on the exact frame (~0.997). fal's audio content filter sometimes rejects a render as "sensitive" (a false positive): re-run it under the next letter.
Kling v3 Pro
fal/fal-ai/kling-video/v3/pro/image-to-video
| Time per render | Middle: slower than H3 Max, faster than Seedance at 1080p |
| Top resolution | 1080p (no resolution field) |
| Look | Holds frame 0 at ~0.996; keeps fine detail and texture better than H3 Max and Flux 3 Max |
| Motion | Simple motion, dependably: smooth (judder ~0.05, like source footage); idles in the opening pose before acting. Not complex physical motion (paper falling like feathers) |
| Inputs | First frame (start_image_url) and end frame; up to 3 @Element references (a character, a product), each a frontal picture plus 1–3 more; no audio input; no seed. Lengths 3–15 s, so a short shot renders ~3–4 s and is trimmed |
| Cost tier | Low to middle; exact prices from describe |
| Leaves | kling-v3-image-to-video |
Where Kling fits: the dependable middle option, an older model that wins no single category. It isn't as fast as H3 Max, can't do complex physical motion like Seedance, and isn't Seedance's photographic 1080p. But it is cheap, reliable on simple motion, holds its first frame, keeps character and product from references, and keeps finer detail than the fast models. Pick it for a simple-motion shot that needs better detail than a draft model gives, without paying for Seedance.
Wan 3 and Flux 3 Max
No leaves yet.
- Wan 3 handles complex motion like Seedance (the best paper physics in one test), but through fal it returned 24 fps padded to 30 with a dup-then-skip cadence (judder 0.78 against 0.05 for the source); decimating and interpolating only reached 0.39.
- Flux 3 Max is fast to iterate on, like H3 Max, and shares its plastic look; Kling keeps finer detail.
TODO(taylor): Wan 3: is there a route without the 30 fps judder now, and when is it worth picking over Seedance for complex motion? Flux 3 Max: endpoint, inputs, and when to pick it over H3 Max for drafts.
Choosing
- Unsure, or not composed yet: H3 Max drafts, then upgrade (above).
- 1080p with realism and photographic grain: Seedance 2.5. H3 can't.
- Complex physical motion: Seedance (or Wan 3, once it has a route), never H3 Max.
- Simple motion, better detail than a draft, without Seedance's cost or wait: Kling v3, the dependable middle option.
- Close-up fine texture (beards, skin, hair): Seedance or Kling. H3 Max only when the face is small on screen.
- Plastic skin on an otherwise good take: grain in post before a re-render
(
guide-prepare-renders-for-delivery). Grain won't fix shine that is in the light. - Audio in: only some models take audio. Seedance r2v (track up to 30 s) and H3 Max
r2v (up to 15 s) take a dialogue track; H3 Max i2v lip-syncs a still to one. Kling and
Seedance i2v take none. See
guide-lipsync-video. - Dialogue scenes: Seedance holds framing and keeps mouths closed on voice-over; H3 Max lip-syncs better and blocks bigger. Render both per shot; picks split about evenly.
- An end frame with dialogue (a walk-out that lands on the next shot): H3 Max r2v.
- Shot length: keep every render to about 15 s or less, on any model; split a longer shot at a pause. H3 Max renders as short as about 1 s; Kling's floor is 3 s, so a short shot renders longer there and is trimmed.
- Comparing takes for smoothness:
python-measure-judderacross takes of one shot.
Avoid
- Plain H3: H3 Max is its fine-tune, so always use Max (plain H3 also needs ≥ 5 s).
- Veo 3.1: rejected on look.
- H3 Max for 1080p or photographic realism, or for complex physical motion.
- Omni Flash, regular H3 r2v, Wan r2v when the layout must hold: no first frame, they redraw the opening.
Related
guide-animate-a-shot (routes: i2v, pinned ends, ref2vid, recast), guide-lipsync-video,
guide-prepare-renders-for-delivery (Topaz, grain).
