Comparison
PixVerse V6 Max vs Other Video Models: How to Choose and Migrate
A developer's cost and capability walkthrough for text-to-video and image-to-video routes, with a practical migration path
На этой странице
What PixVerse V6 Max Actually Exposes
PixVerse V6 Max is not a single endpoint with a flag bolted on. It is a model family that exposes two distinct routes, and the route you pick determines which parameters exist, how long your prompt can be, and how much control you have over the frame.
Route one: text-to-video
The text-to-video route takes a single required input — a prompt of up to 5000 characters — and generates a video from scratch. Because there is no source frame, this route owns the full geometry of the output: aspect_ratio is adjustable across a wide enum (16:9, 4:3, 1:1, 3:4, 9:16, 2:3, 3:2, 21:9), defaulting to 16:9. If your product needs vertical output for a feed, a square cut for a carousel, or an ultrawide banner, this is the route that can produce it.
The long prompt ceiling matters more than it looks. 5000 characters is enough for a structured scene description — camera movement, lighting, subject action, environment, pacing — without aggressive compression. Teams migrating from models with tighter limits often find they can drop their prompt-splitting hacks entirely.
Route two: image-to-video
The image-to-video route animates a supplied first frame. It requires both prompt and image_urls, and the image constraint is strict: exactly one public URL (minItems=1, maxItems=1), JPEG, PNG or WebP, up to 20 MB. The prompt ceiling is lower here — 2048 characters — because the visual content is already established. Your prompt describes motion and atmosphere, not composition.
The important consequence: the output follows the input image's aspect ratio. There is no aspect_ratio parameter on this route at all. If you need a specific shape, you control it beforehand by choosing or cropping the source image.
What is shared
Both routes bill per second and share the same tier structure: resolution (360p, 540p, 720p, 1080p, default 720p), duration (1–15 seconds, default 5), generate_audio (boolean, default False), and seed (0 to 2147483647). Same seed plus same prompt reproduces a similar result, which is the closest thing to determinism you get in video generation and is worth wiring into your test harness.
Both routes are available under the HiAPI.ai unified interface, so the request shape and auth flow are consistent across the family — see the text-to-video route and the image-to-video route for the full parameter schemas.
Text-to-Video vs Image-to-Video: Parameter Differences That Matter
The two routes look similar in a docs table until you check which fields are required and which simply do not exist. That asymmetry is where most integration bugs originate.
Required inputs
Text-to-video requires only prompt. Everything else — resolution, duration, aspect ratio, audio, seed — has a default or is optional. You can fire a request with one field.
Image-to-video requires prompt and image_urls. There is no valid request without a source frame. If your pipeline sometimes has an image and sometimes does not, you are really building two code paths, not one.
Where aspect ratio lives
This is the single most consequential difference. aspect_ratio exists only on text-to-video, with nine enum values and a 16:9 default. On image-to-video it is absent; the output inherits the source image's ratio.
Practically, this means:
- A prompt-first product that also supports image uploads cannot expose one "output shape" dropdown to all users. Image-first jobs get their shape from the upload.
- If you need a guaranteed 9:16 output from an image, validate or crop the image before you send it. Do not expect the API to letterbox or reframe it.
Prompt budget
Text-to-video allows up to 5000 characters; image-to-video allows up to 2048. A shared prompt template that produces long, richly specified prompts will pass on one route and fail validation on the other. Either trim per route or keep two templates.
Shared controls
resolution, duration, generate_audio, and seed behave identically on both routes. duration is an integer from 1 to 15 and is the parameter billing actually multiplies against. generate_audio defaults to False on both — audio is opt-in, not implied. seed is a 32-bit-ish integer range shared by both, so a seed captured from one route is syntactically valid on the other, though it will not reproduce the same visual.
One audio-specific detail worth internalizing: to get spoken lines, you put them in quotes inside the prompt. Audio generation covers background music, sound effects, and dialogue, but dialogue is prompt-driven rather than a separate field.
Decoding the Pricing Tiers: Resolution and Audio Multipliers
PixVerse V6 Max bills per second of output video. Duration is the multiplier everyone remembers, but duration is not where budgets actually diverge. Resolution and audio are the real cost levers, and they compound.
Current per-second rates live on the pricing panel; what matters here is the structure. The official tier grid has eight cells — four resolutions crossed with sound off and sound on:
- 360P · Sound Off
- 540P · Sound Off
- 720P · Sound Off
- 1080P · Sound Off
- 360P · Sound On
- 540P · Sound On
- 720P · Sound On
- 1080P · Sound On
The shape of that grid tells you three things.
Resolution is roughly binary at the top
The meaningful step is between 720P and 1080P, not between the lower tiers. 720P sits at the base per-second rate, while 1080P is roughly double it. Going from 360P to 720P is a smaller proportional jump than going from 720P to 1080P. If you are prototyping, the lower tiers are a cheap place to iterate; if you are shipping to a large screen, the last step is the expensive one.
Audio is a meaningful, not trivial, surcharge
Turning sound on adds cost at every resolution. It is not a rounding error — it is a step change comparable to moving up a resolution tier at some points in the grid. If your product does not need audio, leaving generate_audio at its default of False is a straightforward saving, and the default already does that for you.
The multipliers compound
The headline comparison: 1080P with sound is roughly 4.7x the 360P sound-off rate. That is two independent decisions — maximum resolution and audio on — stacking on top of each other. Neither is unreasonable in isolation; together they define your worst-case per-second cost.
Multiply any tier rate by duration and you have per-job cost. A 15-second 1080P job with audio is 15 times the most expensive tier; a 5-second 360P silent job is 5 times the cheapest. Same model, same prompt, wildly different bill. Budget for the configuration, not the model.
The base price is a baseline, not a quote
The advertised base per-second rate corresponds to the 720P sound-off tier. Any quote you build into a UI should be computed from the tier you actually pin, never from the base rate — otherwise 1080P audio jobs silently overshoot your estimate by a large factor.
Choosing a Configuration for Your Use Case
Model specs tell you what is possible; workload shape tells you what to pick. Here is how to reason about each lever.
Resolution
Start at the default 720p and move only with a reason.
- Feed thumbnails, in-app previews, social loops: 360p or 540p is usually indistinguishable at the display size, and lets you generate far more candidates per budget.
- Standard web playback: 720p is the natural default — it is the base tier, and it is what the model assumes if you say nothing.
- Full-screen or large-display output: 1080p. Accept that you are paying roughly double, and reserve it for the final render rather than the exploration pass.
A common pattern is to explore at low resolution, pick the winning seed and prompt, then re-render the winner at high resolution. Because the same seed and prompt reproduce a similar result, this two-stage approach is coherent rather than wasteful.
Duration
duration runs 1 to 15 seconds and multiplies the per-second rate directly. The default of 5 seconds is a sensible first target: long enough to show motion, short enough to keep iteration cheap. Avoid a 15-second max-duration default in any batch job — it is the single fastest way to inflate spend without improving output quality.
Audio
Default is off, and for good reason. Generate audio only when the output needs it: narrated explainers, ambience-driven scenes, anything with dialogue. If you do need dialogue, remember that spoken lines belong in quotes in the prompt — that is the mechanism, and it costs prompt characters on both routes.
Prompt length
Text-to-video gives you up to 5000 characters; image-to-video up to 2048. Do not pad to the ceiling. Use the extra room on text-to-video for camera and lighting direction; on image-to-video, spend your budget on motion verbs and atmosphere, since the composition is already fixed by the frame.
Route choice
Ask one question: does the system already have a picture of the first moment?
- If yes — product shots, avatars, existing photography, a storyboard frame — use image-to-video. You get more visual control and a shorter prompt to maintain.
- If no — you are generating the scene itself — use text-to-video, and use
aspect_ratiodeliberately rather than accepting 16:9 by inertia.
Seed
Treat seed as a testing tool, not a production constant. Pin it during prompt iteration so you can tell whether a change in output came from the prompt or from sampling noise. In production, vary it — or expose it, if your users want to re-roll a near-miss.
Setting Up a Cost-Aware Integration
Per-second billing with three compounding levers is exactly the kind of system where defaults quietly become your budget. The fix is structural: pin the cost levers explicitly at the boundary of your application, and never let an unspecified default decide them.
Pin the three levers in one place
Define a single configuration object per workload type — for example preview, standard, and delivery — each specifying resolution, duration, and generate_audio explicitly. Pass that object through to your request builder instead of relying on omission. The model's defaults (720p, 5s, audio off) are a reasonable baseline, but a baseline you did not choose is a baseline you cannot defend when the invoice arrives.
Validate before you send
Cheap client-side checks catch the expensive mistakes:
durationmust be an integer within 1–15. Clamp it, do not hope.resolutionmust be one of360p,540p,720p,1080p. Reject anything else rather than falling back silently.aspect_ratioexists only on text-to-video. If your request builder is shared, strip or reject it on image-to-video rather than letting an unknown field ride along.image_urlsmust be a single-element array pointing to a public URL in JPEG, PNG or WebP, under 20 MB. Check reachability before submitting; a broken URL is a wasted round trip.- Prompt length: enforce 5000 characters on text-to-video and 2048 on image-to-video, per route.
Estimate cost at request-build time
Compute an estimate from duration times the tier rate for the exact (resolution, audio) pair you selected, and log it alongside the job ID. This gives you per-job cost attribution without querying anything after the fact, and it makes anomalies visible — a sudden jump in average estimated cost per job usually means a duration or resolution default drifted somewhere in the code.
Guard the fan-out
Cost-aware integration is mostly about concurrency policy. Cap the number of simultaneous high-resolution jobs; queue or downgrade the rest. If you run an exploration phase, hold it to the low tiers and only promote winners to delivery settings. That one rule usually dominates every other optimization.
Use the right route for your pipeline shape
Prompt-first pipelines should live entirely on text-to-video; image-first pipelines should live entirely on image-to-video. If you need both, model them as two workloads with two configs and two prompt templates, not one function with a branch — the required fields and prompt limits genuinely differ.
The route schemas are the authoritative reference for field names and enums: text-to-video and image-to-video.
Migrating from Another Video Model
Migration to PixVerse V6 Max is less about porting code and more about re-mapping three axes: prompt, resolution, and audio. Do it in a fixed sequence so you can attribute regressions.
Step 1: Classify your pipeline
Before touching fields, decide what your system fundamentally is. Does every job start with a prompt, or does some part of your product carry a source image? This determines your target route and is the one decision that cascades into everything else. Write it down per endpoint, not per service.
Step 2: Map the fields
Build an explicit mapping table from your current request shape to the target route. The mandatory ones:
- Your prompt field →
prompt, with length limits enforced at 5000 (text-to-video) or 2048 (image-to-video). - Your length control →
duration, integer 1–15, default 5. - Your quality control →
resolution, enum360p/540p/720p/1080p, default 720p. - Your audio flag →
generate_audio, boolean, defaultFalse. - Your first-frame / init-image field →
image_urls, a single-element array, image-to-video only. - Your seed →
seed, integer 0 to 2147483647.
Anything in your current request with no counterpart should be dropped loudly, not passed through silently. Unknown-field tolerance is a migration trap: it hides the fact that a feature you depend on does not exist on the new route.
Step 3: Rewrite prompts
Prompts rarely transfer verbatim. Do a pass over your templates and:
- Expand where you were compressing unnaturally — 5000 characters on text-to-video is generous, so restore detail you had to cut.
- Move spoken dialogue into quotes. If your old integration had a dedicated narration field, that concept does not exist here; the mechanism is quoted lines inside the prompt.
- On image-to-video, rewrite from composition language to motion language. The frame already establishes what is in the shot; your prompt should describe how it moves and what the atmosphere feels like, within 2048 characters.
Step 4: Handle aspect ratio correctly
This is the migration step most likely to produce a visible regression.
- Text-to-video: map your old size control onto
aspect_ratio(16:9, 4:3, 1:1, 3:4, 9:16, 2:3, 3:2, 21:9). If you were passing pixel dimensions, convert them to the nearest ratio and record the mapping. - Image-to-video: there is no
aspect_ratio. Output follows the source image. Audit your image sources — including user uploads — and normalize them before generation if you need a guaranteed shape.
Step 5: Re-baseline your cost model
Per-second billing with resolution and audio tiers will not match whatever your previous model charged. Recompute your per-job estimates from the official per-second grid at the durations you actually use: with sound off, $0.0305 / $0.0458 / $0.0559 / $0.1118 at 360P / 540P / 720P / 1080P; with sound on, $0.0458 / $0.0559 / $0.0762 / $0.1423. Update any user-facing quota or credit conversion. Do this before launch, not after the first invoice.
Step 6: Validate output
Run a golden set of prompts and images through both routes. Check that files decode, durations match what you requested, aspect ratios match expectations, and audio is present only where you asked for it. Pin seed during this phase so differences reflect your changes rather than sampling variance.
Step 7: Cut over gradually
Route a small percentage of traffic to the new integration with a per-job cost log attached, compare against your baseline, then widen. Because both routes sit behind the same unified interface on HiAPI.ai, cutover is a configuration change rather than a new vendor integration.
Common Pitfalls and How to Avoid Them
These are the failure modes that show up repeatedly once teams go past the happy path.
Expecting aspect ratio control on image-to-video
The most common mistake. Teams build one request builder for both routes, include an aspect_ratio field, and are surprised when vertical output does not appear. aspect_ratio is a text-to-video parameter. On image-to-video, the output follows the source image's ratio, full stop. Fix: normalize or crop source images first, and strip aspect_ratio from image-to-video requests entirely rather than letting it be ignored.
Assuming audio is on by default
generate_audio defaults to False on both routes. Teams that ship a demo and hear silence conclude the audio feature is broken. It is not — it is opt-in. Fix: set the flag explicitly in any workload that needs sound, and remember dialogue requires quoted lines in the prompt rather than a separate input.
Exceeding the prompt limit on the wrong route
Shared templates built for the 5000-character text-to-video ceiling will fail on image-to-video's 2048-character limit. Fix: enforce length per route at validation time, and keep the image-to-video template focused on motion and atmosphere so it naturally stays shorter.
Passing multiple images
image_urls accepts exactly one URL. Arrays of two or more are invalid. Fix: if your product supports multi-image upload, pick the first frame deterministically (first upload, user-selected, or a documented rule) and send one URL.
Unreachable or invalid image URLs
The image must be publicly accessible, JPEG/PNG/WebP, and under 20 MB. Private buckets, expiring signed links, and HEIC uploads are all common causes of failed jobs. Fix: verify reachability and content type before submitting, and re-host where necessary.
Relying on defaults for cost-sensitive batch jobs
The defaults — 720p, 5 seconds, audio off — are safe individually but are not a budget policy. A batch job that never pins duration or resolution cannot be reasoned about when costs rise. Fix: pin all three cost levers explicitly for batch workloads and log the estimate per job.
Treating a seed as a guarantee
Same seed plus same prompt reproduces a similar result — it is a stability aid, not a deterministic render. Fix: use seeds for A/B iteration during prompt tuning, and do not build product features that promise byte-identical regeneration.
Out-of-range durations
duration is an integer from 1 to 15. Fractional values, zero, and anything above 15 are invalid. Fix: clamp at the boundary and surface the clamped value in your UI so users know what they will get.
Avoid these eight and the integration stays boring, which is exactly what a per-second-billed video pipeline should be. Start with the text-to-video route or the image-to-video route depending on whether your pipeline is prompt-first or image-first.
Похожие статьи
Все статьи →-
Comparison
gpt-6.2-sol vs. comparable models: How developers should choose and migrate
Make a side-by-side decision across three dimensions: capability boundaries, pricing structure, and migration cost
-
Comparison
gpt-6.2-astra Side-by-Side Comparison and Selection Guide: How Developers Can Decide Whether It Is Worth Integrating
-
Comparison
Tripo H3.1 Side-by-Side Comparison and Selection Guide: How to Choose Among Three 3D Generation Paths and How to Save
A developer-oriented trade-off between pricing and capability, starting from input form, geometric refinement, and PBR/texture switches