Every generation in DOTV consumes credits: the model, resolution and length you choose directly decide how many it costs. This section puts "which models are there" and "how is the money worked out" side by side, because they were always two sides of the same thing: where you pick the model, what you get right now, what you are charged by, how the price is worked out, and the estimate versus what you are actually charged.
Select a node and the input bar floats up. The first control at its bottom-left is the model. Each kind of node has its own list (image models never show up on a video node).
Change the model and the resolution and duration steps beside it change with it — providers support different ones, and picking a step the other side does not accept buys you nothing. So after switching models, take a look at that Aspect · Resolution · Duration button.
| Used for | Model | Notes |
|---|---|---|
| Image | Gemini Image (Nano Banana 2) | The default for both generating and editing |
| Image | Gemini Image Pro (Nano Banana Pro) | First tier for text rendering and complex layouts; up to 4K |
| Image | Seedream 4.5 Image | Strong on Chinese scenes and poster typography; the volume option |
| Image | Seedream 5.0 Pro / Lite Image | Pro is better at multi-image fusion and editing, Lite is faster and cheaper; synchronous — one image in a dozen seconds |
| Image | FLUX.2 Pro Image | The photo-grade option |
| Image | Midjourney | Submit then poll (a minute or two); one image per run, auto-upscaled from the grid |
| Image | GPT Image 2.5 | OpenAI image model, steady on complex layouts; the spec popover adds a Quality row (only this model has it) next to resolution (1K / 2K / 4K) and count, and it takes up to 10 reference images |
| Video | Seedance 2.0 mini Video | 480P / 720P, 4–15 seconds, and duration can be left on Auto (the model decides) |
| Video | Seedance 2.5 Video | Adds 1080p, native audio-sync, and up to 50 omnimodal references |
| Video | MiniMax H3 Video (2K) | 2K only, and no Auto duration |
| Video | MiniMax H3 Max Video | The same generation's speed option: 480P / 768P, faster and cheaper |
| Video | Kling 3.0 Standard / Pro / Turbo / 4K Video | One tier per endpoint: Standard is cheapest, Turbo is fast with a native soundtrack (can't be muted), 4K renders 4K in one step |
| Video | Veo 3.1 / Fast / Lite Video | Audio-sync is the selling point; duration only comes in 4 / 6 / 8 s; Lite is the cheapest audio-synced option on the site |
| Video | Vidu Q3 Pro / Turbo Video | Audio-sync included in the price; 1–16 s, 540p–1080p |
| Video | Wan 3.0 / Prime Video | Omni-reference architecture; Prime is the speed tier — same abilities, faster |
| Video | Kling O3 Standard / Pro / 4K Video | The only family that takes video in, video out: multi-image references, video references (the source clip for extends / reshoots / 3D previz really reaches the model), instruction-based editing |
| Video | Happy Horse 1.1 Video | Arena leader for both text-to-video and image-to-video; native audio-sync with phoneme-level lip-sync in several languages, soundtrack can't be muted |
| Text | DoTV LM 2.0 | The default, a product-layer model (backed by deepseek-v4.1-flash with product-level prompt constraints; with audio or video attached it runs on Doubao Seed 2.0 Lite Omni instead) |
| Text | DeepSeek V4.1 Flash | The bare model: what you pick is what gets called |
| Text | GLM-5.3 Flash | Bare vision model; web search runs through Aihubmix :surfing |
| Text | Doubao Seed 2.1 Turbo | Bare model; the volume tier — fast output at a low price |
| Text | Claude Opus 5 | Top reasoning tier, and the priciest — one to two orders of magnitude above its neighbours |
| Speech | Minimax-speech-2.8-hd / turbo | hd sounds better, turbo costs less |
| Speech | Qwen3-TTS-Flash / Doubao-TTS-2.0 | Alternatives |
| Music & SFX | ACE-Step Music / ElevenLabs Music / Doubao Music (GenBGM) / Mureka / ElevenLabs SFX / Doubao Seed-Audio | Scores and effects, each from its own prompt |
| Video understanding | Doubao Seed 2.0 Lite Omni (with Audio) | Only appears in the list when you are writing a script from a video (see 4.7). It hears sound — dialogue and music need it |
| Video understanding | Doubao Seed 2.1 Pro Vision | Stronger on the picture, but deaf: give it the work that is only about what's on screen |
Resolution, aspect, duration and the rest are covered in their own chapters (images in 3.2, video in 4.2, speech in 5.2).
Claude Opus 5 is the strongest text-reasoning option here, and it is clearly more expensive. Use it for complex creation and multiple constraints: turning a long script into coherent shots, organizing character relationships across scenes, revising a plan against brand, duration, platform and tone at once, or pulling structure back together after many rounds of edits.
For everyday short copy, simple renaming or polishing a single prompt, do not make it the default. Start with an ordinary text model to settle structure and material, then use the stronger reasoning model for the harder organization; that usually saves credits. Pricing changes with content length, so trust the price shown in the input bar before submitting.
Different jobs are priced on different things — this is the answer to "why do two runs both called video differ tenfold":
| What you're doing | What it's priced by |
|---|---|
| Image | Per image (Midjourney is one image per run, with no count to set) |
| Video | Per second × resolution (higher costs more, and not linearly) |
| Speech | Per character to be read out |
| Music & SFX | A flat price per track, or by generated length (depends on the model) |
| Video soundtrack | Per second — the surcharge for turning sound on for a render (models with a native soundtrack don't have this line) |
| Text | By content length (a longer prompt costs more) |
| Video upscale | Per second × target resolution; with slow motion on, by the stretched duration |
| Image upscale | Banded by output area — a small image at 4x can cost less than a big one at 2x |
| Image editing | Expand / regenerate / erase / cut out background: a flat price per image (cutting out is markedly cheaper — it makes no new pixels) |
| Remove subtitles | Per second; the auto-detect band costs a little more than boxing it yourself |
| Audio separation | Per second |
| Voice cloning | Per run (one run covers building the voice plus the sample it reads back) |
| Video understanding | By how many frames were sampled (a higher sampling rate costs more) |
| Face detection | Per run (one image, one run) |
The provider's cost × our markup (2.2–2.7× depending on the model), rounded up to whole credits. 1 credit = ¥0.04.
Two rules follow from that:
The number at the bottom-right of the input bar is an estimate: based on the model, resolution, duration and count you have selected right now, and it moves the moment you change one of them.
It and the charge you actually pay follow the same rules, so you will never see 18 on screen and 42 on the bill. There is exactly one case where they differ: when the model decides a parameter for itself (duration set to Auto, say), the estimate falls back to the default value.
If your balance is short the submission is rejected outright — it tells you what this run needs and what you have — rather than failing halfway through. Failed generations are refunded automatically. All of it is recorded in your credit history, see 14.3.