December 5, 2025 (August 9, 2026)
Table of contents
Video Models
Endpoint Compatibility
| Model | create | create-frames | create-fusion | motion-control |
|---|---|---|---|---|
v6 (default) | ✓ | ✓ | — | — |
v5.6 | ✓ | ✓ | ✓ | ✓ |
pixverse-c1 | ✓ | ✓ | ✓ | — |
seedance-2.5 | ✓ | ✓ | ✓ (+10 ref videos, +10 ref audios) | — |
seedance-2.0 | ✓ | ✓ | ✓ (+3 ref videos, +3 ref audios) | — |
seedance-2.0-fast | ✓ | ✓ | ✓ (+3 ref videos, +3 ref audios) | — |
seedance-2.0-mini | ✓ | ✓ | ✓ (+3 ref videos, +3 ref audios) | — |
kling-o3 | ✓ | ✓ | ✓ | — |
kling-v3 | ✓ | ✓ | — | — |
grok-imagine | ✓ | — | — | — |
grok-imagine-1.5 | ✓ (i2v only) | — | — | — |
veo-3.1-lite | ✓ | ✓ | — | — |
veo-3.1-standard | ✓ | ✓ | — | — |
veo-3.1-fast | ✓ | ✓ | — | — |
sora-2 | ✓ | — | — | — |
sora-2-pro | ✓ | — | — | — |
happyhorse-1.0 | ✓ | — | — | — |
Fusion notation: v5 uses @pic1/@pic2/@pic3, all other fusion-capable models use @image1…@imageN (mapped positionally to frame_1_path…frame_N_path). The Seedance fusion family additionally supports @video1…@videoN for reference videos (video_1_path…video_N_path) and @audio1…@audioN for reference audios (audio_1_path…audio_N_path) — PixVerse’s omni mode. seedance-2.5 takes up to 10 of each and 30 images, the 2.0 family 3 of each and 9 images.
grok-imagine-1.5 is image-to-video only — it requires first_frame_path and rejects text-to-video (no-image) requests. The original grok-imagine supports both text-to-video and image-to-video.
Extend, upscale, modify, lipsync
Upscale, modify, and lipsync are native-PixVerse only. Extend supports v6 and the third-party grok-imagine model.
extend also accepts grok-imagine (480p/720p, 2-10s, native audio).
v5 family (legacy)
v5 carries legacy-only modes — multi-frame create-transition, lipsync, and fusion with the original @pic1/@pic2/@pic3 notation.
| Endpoint | v5 | v5.5 | v5.6 | v5-fast |
|---|---|---|---|---|
| create | ✓ | ✓ | ✓ | ✓ |
| create-frames | ✓ | ✓ | ✓ | — |
| create-transition (2-frame) | ✓ | ✓ | ✓ | — |
| create-transition (3+ frame) | ✓ | — | — | — |
| create-fusion | ✓ | — | ✓ | — |
| extend | — | — | — | — |
| modify | — | ✓ | — | — |
| lipsync | ✓ | — | — | — |
| upscale | ✓ | ✓ | ✓ | ✓ |
v5 accepts both @image1…@imageN (unified) and the legacy @pic1/@pic2/@pic3 synonyms for backward compatibility.
Quality, Duration, Aspect Ratio
| Model | Qualities | Durations | Aspect Ratios | Max ref (fusion) |
|---|---|---|---|---|
v6 | 360p, 540p, 720p (default), 1080p | 1-15s | 16:9, 9:16, 1:1, 4:3, 3:4 | — |
v5.6 | 360p, 540p (default), 720p, 1080p | 1-10s (1080p max 8) | 16:9, 9:16, 1:1, 4:3, 3:4 | 7 imgs |
v5.5 | 360p, 540p (default), 720p, 1080p | 1-10s (1080p max 8) | 16:9, 9:16, 1:1, 4:3, 3:4 | — |
v5 | 360p, 540p (default), 720p, 1080p | 1-10s (1080p max 8) | 16:9, 9:16, 1:1, 4:3, 3:4 | 3 imgs |
v5-fast | 360p, 540p (default), 720p, 1080p | 1-10s (1080p max 8) | 16:9, 9:16, 1:1, 4:3, 3:4 | — |
pixverse-c1 | 360p, 540p, 720p, 1080p | 1-15s | 16:9, 4:3, 1:1, 3:4, 9:16, 3:2, 2:3 | 7 imgs |
seedance-2.5 | 480p, 720p | 4-30s | 16:9, 4:3, 1:1, 3:4, 9:16, 21:9 | 30 imgs + 10 videos + 10 audios (50 total) |
seedance-2.0 | 480p, 720p, 1080p, 2160p | 4-15s | 16:9, 4:3, 1:1, 3:4, 9:16, 21:9 | 9 imgs + 3 videos + 3 audios |
seedance-2.0-fast | 480p, 720p | 4-15s | 16:9, 4:3, 1:1, 3:4, 9:16, 21:9 | 9 imgs + 3 videos + 3 audios |
seedance-2.0-mini | 480p, 720p | 4-15s | 16:9, 4:3, 1:1, 3:4, 9:16, 21:9 | 9 imgs + 3 videos + 3 audios |
kling-o3 | 720p (Std), 1080p (Pro) | 3-15s | 16:9, 1:1, 9:16 | 7 imgs |
kling-v3 | 720p (Std), 1080p (Pro) | 3-15s | 16:9, 1:1, 9:16 | — |
grok-imagine | 480p, 720p | 1-15s | 16:9, 4:3, 1:1, 3:4, 9:16, 3:2, 2:3 | — |
grok-imagine-1.5 | 480p, 720p | 1-15s | i2v only (from image) | — |
veo-3.1-lite | 720p, 1080p | 4, 6, 8 | 16:9, 9:16 | — |
veo-3.1-standard | 720p, 1080p, 2160p | 4, 6, 8 | 16:9, 9:16 | — |
veo-3.1-fast | 720p, 1080p, 2160p | 4, 6, 8 | 16:9, 9:16 | — |
sora-2 | 720p | 4, 8, 12 | 16:9, 9:16 | — |
sora-2-pro | 720p, 1080p | 4, 8, 12 | 16:9, 9:16 | — |
happyhorse-1.0 | 720p, 1080p | 3-15s | 16:9, 9:16, 1:1, 4:3, 3:4 | — |
aspect_ratiois required fort2vandfusion, not accepted fori2vortransition(derived from image).- For
kling-o3/kling-v3:quality: 720proutes to Std,quality: 1080proutes to Pro. - For
veo-3.1-standard/veo-3.1-fast:quality: 1080prequiresduration: 8. -
Seedance fusion (omni mode) is the only path that accepts reference videos and audios. Images go in
frame_1_path…frame_N_path, videos invideo_1_path…video_N_path(@video1…@videoN), and audios inaudio_1_path…audio_N_path(@audio1…@audioN). At least one image or reference video is required. Every reference must be ≤50 MB for video (24-60 fps, ≤6000×6000 px) and ≤15 MB for audio.seedance-2.5— 30 images, 10 reference videos, 10 reference audios, 50 references in total. Each clip 1.8-30.2s, summed ≤30.2s per kind.seedance-2.0family — 9 images, 3 reference videos, 3 reference audios. Each clip 2-15s, summed ≤15s per kind.
Durations are summed separately for video and for audio.
Audio
| Model | audio |
|---|---|
v6 | toggle |
v5.6 | toggle |
v5.5 | toggle |
v5 | — (use lip_sync_tts_prompt + sound_effect_prompt) |
v5-fast | — |
pixverse-c1 | toggle |
seedance-2.5 | rejected (native audio, always on) |
seedance-2.0 | toggle |
seedance-2.0-fast | toggle |
seedance-2.0-mini | toggle |
kling-o3 | toggle |
kling-v3 | toggle |
grok-imagine | rejected |
grok-imagine-1.5 | rejected |
veo-3.1-lite | rejected |
veo-3.1-standard | always on |
veo-3.1-fast | always on |
sora-2 | rejected |
sora-2-pro | rejected |
happyhorse-1.0 | always on |
toggle— acceptaudio: true/false.always on— audio generated automatically;audio: falseis rejected.rejected—audioparameter is not accepted (content has no audio track or audio is handled internally).
Native PixVerse — extra flags
multi_shot, preview_mode, off_peak_mode, and seed are supported only on native PixVerse models. Third-party models reject them.
| Model | multi_shot | preview_mode | off_peak_mode | seed |
|---|---|---|---|---|
v6 | ✓ | ✓ | ✓ | ✓ |
v5.6 | — | ✓ | ✓ | ✓ |
v5.5 | — | ✓ | ✓ | ✓ |
v5 | — | ✓ | ✓ | ✓ |
v5-fast | — | ✓ | ✓ | ✓ |
pixverse-c1 | — | ✓ | ✓ | ✓ |
Image Models
All image models share the same endpoints: create, list, get, delete.
| Model | Qualities | Max Refs | Est. Time |
|---|---|---|---|
qwen-image (default) | 720p, 1080p | 3 | ~3s |
nano-banana | 1080p | 3 | ~10s |
nano-banana-2 | 512p, 1080p, 1440p, 2160p | 9 | ~30s |
nano-banana-2-lite | 1080p | 14 | ~7s |
nano-banana-pro | 1080p, 1440p, 2160p | 9 | ~60s |
seedream-4.0 | 1080p, 1440p, 2160p | 6 | ~10s |
seedream-4.5 | 1440p, 2160p | 6 | ~15s |
seedream-5.0-pro | 1080p, 1440p | 10 | ~30s |
seedream-5.0-lite | 1440p, 1800p, 2160p | 6 | ~30s |
kling-3.0 | 1080p, 1440p | 1 | ~15s |
kling-o3 | 1080p, 1440p, 2160p | 1 | ~20s |
gpt-image-2.0 | 1080p, 1440p, 2160p | 9 | ~30s |
create_count: 1-4 (default 1).detail_level(gpt-image-2.0only, required):low,medium,high. Rejected for all other models. Affects credit cost (low = 0.5×, medium = 1×, high = 2× of the per-quality base).
Aspect ratios
Each model accepts its own list. If aspect_ratio is omitted, the default (first column) is used. Passing a value not in the model’s row is rejected with 400.
| Models | Default | Accepted aspect_ratio values |
|---|---|---|
nano-banana, nano-banana-2, nano-banana-2-lite, nano-banana-pro, seedream-4.0, seedream-4.5, seedream-5.0-pro, seedream-5.0-lite | auto | auto, 1:1, 16:9, 9:16, 4:3, 3:4, 5:4, 4:5, 3:2, 2:3, 21:9 |
qwen-image | 1:1 | 1:1, 16:9, 9:16, 4:3, 3:4, 5:4, 4:5, 3:2, 2:3, 21:9 |
kling-o3, kling-3.0 | 1:1 | 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9 |
gpt-image-2.0 | 1:1 | 1:1, 1:2, 4:3, 3:4, 3:2, 2:3, 16:9, 9:16, 2:1, 21:9 |
Unlimited Image Generation (Relax Mode)
Pro+ subscription plans include unlimited image generation in Relax Mode for select models:
| Plan | Price | Unlimited Models |
|---|---|---|
| Pro | $30/m | qwen-image, nano-banana-2-lite |
| Premium | $60/m | qwen-image, nano-banana, nano-banana-2, nano-banana-2-lite, seedream-4.0 |
| Ultra | $199/m | qwen-image, nano-banana, nano-banana-2, nano-banana-2-lite, nano-banana-pro, seedream-4.0, seedream-4.5, seedream-5.0-lite, seedream-5.0-pro, kling-3.0, kling-o3, gpt-image-2.0 (all 12) |
Relax Mode is unmetered, not unthrottled. These models cost zero credits on the listed plans and nothing rejects a request for being too frequent, but PixVerse applies a daily volume threshold. Past it the account is moved to a slow queue rather than refused — generations still succeed and still cost nothing, they just take far longer to come back. It is visible in the response, where queue_data.slow_queue turns true, queue_data.reason reads daily_limit_exceeded, and queue_count gives the number of tasks ahead.
The threshold applies per model, not per account. An account throttled on one model keeps generating at full speed on every other one, and the busiest accounts clear 500–700 images a day across models combined. We have seen it engage on nano-banana-2, nano-banana-pro and gpt-image-2.0, most often on nano-banana-2. The point at which it engages is not fixed and varies between accounts, but as a rough guide expect a few hundred images per model per day — around 300 for nano-banana-pro and around 500 for gpt-image-2.0. We have not seen it on the remaining models, which does not mean they are unlimited, only that nobody has pushed them hard enough for it to show. The allowance resets at approximately 00:00 UTC, so a throttled account is clear again at the start of the next UTC day. PixVerse moves these thresholds without notice, so treat them as rough capacity guidance rather than a guarantee.
Sustained volume needs more than one account, because each account carries its own allowance and POST /videos/create and POST /images/create spread work across every account you have configured. Adding accounts alone is not a guarantee though — we have seen a multi-account pool throttled on the same model at the same time, because every account was working through the same daily allowance on the same model. Once a pool is in that state more accounts only buy one more allowance each, and none of them are fast.
Falling back to another model is usually the better move, because the allowances are independent per model. An account wedged on nano-banana-2 still generates on nano-banana-pro, seedream-4.5 or gpt-image-2.0 at full speed. Decide in advance which models are acceptable for your use case, watch for queue_data.slow_queue turning true on the image or video response, and move down that list rather than waiting out the queue.
Music Models
All music models share the same endpoints: generate, list, get. Generation is asynchronous — poll audio_status_final or pass a replyUrl. All models generate at an automatic duration (typically 2-5 minutes).
| Model | Provider | Prompt max | Custom lyrics | Reference images | Credits |
|---|---|---|---|---|---|
music-2.6 (default) | MiniMax | 2,000 | ≤ 3,500 chars | — | 40 |
music-v1 | ElevenLabs | 4,000 | ≤ 3,500 chars | — | 150 |
lyria-3-pro-preview | 5,000 | — | up to 10 | 20 |
Output modes
The output is derived from instrumental and lyrics — there is no auto_lyrics parameter.
| Request | Result |
|---|---|
instrumental: true | instrumental, no vocals |
lyrics provided | vocals sung from your lyrics (music-2.6 / music-v1 only) |
| neither | vocals with lyrics written by the model |
lyricsis rejected onlyria-3-pro-previewand cannot be combined withinstrumental.image_path_1…image_path_10are accepted bylyria-3-pro-previewonly, provided sequentially.
Status values
| audio_status | audio_status_name | audio_status_final |
|---|---|---|
| 1 | COMPLETED | true |
| 5 | QUEUED | false |
| 8 | FAILED | true |
| 10 | GENERATING | false |
A failed track (audio_status 8) carries fail_code and a human-readable fail_reason. These are usually transient backend errors — re-submitting the same request often succeeds.
Lyria 3 Pro is also available through Flow Music at a lower price with many more features — cover and restyle, lyrics-adjust remix, extend and replace, and audio effects.
Text-to-Speech Models
All speech models share the same endpoints: generate, list, get, plus voices and models for discovery. Generation is asynchronous — poll audio_status_final or pass a replyUrl. Credits are billed per character — credits = ceil(chars / N).
| Model | Provider | Voice settings | Max chars | N (chars/credit) |
|---|---|---|---|---|
eleven-multilingual-v2 | ElevenLabs | stability, similarity_boost, speed, style | 10,000 | 50 |
eleven-v3 | ElevenLabs | stability, similarity_boost, speed + audio tags | 5,000 | 50 |
eleven-turbo-v2.5 | ElevenLabs | stability, similarity_boost, speed | 40,000 | 100 |
speech-2.8-hd (default) | MiniMax | speed, volume, pitch, emotion | 10,000 | 50 |
speech-2.8-turbo | MiniMax | speed, volume, pitch, emotion | 10,000 | 100 |
Every request needs a voice_id from GET speech/voices (filtered by model and language). The provider and the paired provider_voice_id are derived automatically.
Voice settings
The two providers take different settings, enforced per family (a setting from the wrong family is rejected with 400):
- MiniMax —
speed(0.5-2),volume(0-10),pitch(-12-12),emotion(auto,happy,sad,angry,fearful,disgusted,surprised,neutral,calm). - ElevenLabs —
stability(0-1),similarity_boost(0-1),speed(0.7-1.2).style(0-1) anduse_speaker_boostare accepted byeleven-multilingual-v2only.
Expressive control
| Model | Control |
|---|---|
eleven-v3 | inline audio tags in the text — [whispers], [excited], [shouts], [sighs], [laughs], [curious], … Works best with longer, sentence-level text. |
speech-2.8-hd / speech-2.8-turbo | the emotion voice setting (a natural, subtle coloring). |
eleven-multilingual-v2 | the style voice setting. |
eleven-v3 is the most expressive model. The other ElevenLabs models read audio tags literally — use them only with eleven-v3.
Languages
language_code is validated per model against the live models catalog (MiniMax and the ElevenLabs v2/turbo models support 30-40 languages). eleven-v3 auto-detects the language and rejects language_code with 400.
Status values
| audio_status | audio_status_name | audio_status_final |
|---|---|---|
| 1 | COMPLETED | true |
| 5 | QUEUED | false |
| 8 | FAILED | true |
| 10 | GENERATING | false |
A failed job (audio_status 8) carries fail_code and a human-readable fail_reason. These are usually transient backend errors — re-submitting the same request often succeeds.