ZenCreator MCP Docs
/
Connect
View as Markdown Open in ChatGPT Open in Claude

Generation tools & models#

This page is the catalog of generation tools — the values you pass as the tool_name argument to zencreator_create_task — and the models each generation tool offers.

Keep two layers distinct (see Concepts):

  • MCP tools are the zencreator_* tools your AI client sees. zencreator_create_task is one of them.
  • Generation tools are the tool_name values you pass into zencreator_create_task — e.g. by_prompt, image_editor, videogen, faceswap, upscaler, lipsync. Each generation tool exposes its own set of models (the model field).

Plainly: zencreator_create_task is an MCP tool; by_prompt is a generation tool you pass to it; FLUX_KLEIN_NSFW is a model you pass to by_prompt.

How to use this catalog#

  1. Pick a generation tool below for the task (text→image, edit, video, faceswap, …).
  2. Pick a model from that tool's list. Models marked (trusted) require a trusted account for NSFW use.
  3. Call zencreator_get_tool_schema for the exact input/output JSON Schema, a prompt-writing guide, and model-selection guidance before submitting.
  4. For per-model prompt conventions (Seedream prose vs. Qwen layout-first vs. Wan structured blocks, text-in-image rules, NSFW phrasing), call zencreator_get_model_prompt_guide.

NSFW and trusted gating#

ZenCreator supports uncensored / NSFW generation. Two account flags are involved:

  • nsfw_allowed — the "Show 18+ content" toggle: whether 18+ content is shown. Not a generation gate — never stop a generation because it is false.
  • is_trusted — whether the account has Trusted Status. It unlocks ZenCreator's extended capabilities (uncensored NSFW generation, 18+ templates and LoRAs, Face Swap tools, more flexible generation) and gates the (trusted) models below plus the trusted-only tools flux_klein_lora and text_to_video. Trusted status is granted automatically after the account's first successful payment (buying any credit pack in Billing) and is permanent; until then, submitting a trusted-only task fails and wastes credits.

Check both proactively with zencreator_get_me before an NSFW workflow. See Workflows for the full NSFW preflight.

The optional zencreator_craft_prompt sidecar (present on deployments that configure it) can author NSFW / model-specific prompts when the orchestrator itself declines to — it spends no ZenCreator credits.

Choosing & comparing#

Credit cost varies by generation tool, model, resolution, and duration — and it is backend-driven, so it drifts. Never hardcode credit numbers.

  • zencreator_compare_prices — shop across all models of a generation tool, cheapest-first (sweeps resolutions when relevant). Use this whenever the user wants the cheapest option.
  • zencreator_estimate_price — get the exact credit cost of one candidate input. Call this and state the cost to the user before submitting.

Image generation tools#

by_prompt — Text-to-image#

Generate an image from a text prompt, with no input image. The entry point for creation: use it when the user has no source image and describes the picture in words — building a character or content from scratch, concepts, backgrounds, NSFW from a description, quick drafts and final hero shots. Supports fast/quality modes, batches, aspect ratios and body-shape LoRAs.

Three content groups, pick by what you need:

  • Top cloud models, censored (GENERAL_NSFW, NANO_BANANA, NANO_BANANA_2_1, GPT Image 2.5, IDEOGRAM_4_5) — highest quality and realism, but block explicit content.
  • Uncensored but not built for porn (WAN_2_7_IMAGE, QWEN_IMAGE, SEEDREAM_5, SEEDREAM_5_PRO) — high quality and won't block, but won't create explicit content from scratch; they accurately transform NSFW references you provide.
  • Local, built for explicit NSFW (FLUX_KLEIN_NSFW, Z_IMAGE_SPICY) — slightly lower quality and more artifacts, but real explicit capability.

Models:

  • GENERAL_NSFW (default, trusted) — General-purpose workhorse with a good quality/speed/price balance and strong facial likeness; the NSFW version is uncensored. Does not produce explicit anatomy from text alone (it covers it up). Older model — occasional hand/limb artifacts.
  • GENERAL_SFW — The same pipeline, SFW only.
  • WAN — Legacy WAN image model; prefer WAN_2_7_IMAGE.
  • WAN_2_7_IMAGE / WAN_2_7_IMAGE_PRO — Modern model with higher quality and detail (Pro = top consistency). Renders bodies, scenes and composition more aesthetically with fewer hand/limb artifacts. Weaker at in-image text. Uncensored, but transforms your NSFW references rather than inventing explicit content.
  • QWEN_IMAGE / QWEN_IMAGE_PRO — Aesthetic results with good facial likeness and few artifacts; great for stylized / illustrative / anime subjects. Pro adds realism. Uncensored; transforms NSFW references.
  • SEEDREAM_5 — Newer generation: better prompt understanding, stronger stylization, better likeness, fewer artifacts. Uncensored; transforms NSFW references.
  • SEEDREAM_5_PRO — The newest Seedream flagship: best prompt understanding and realism of the line. The only model here with all 15 aspect ratios (adds 1:2, 2:1, 1:3, 3:1, 4:5, 5:4, 9:21) and a tier switch: 1K = 2 credits, 2K (default) = 3 credits. Uncensored; transforms NSFW references.
  • KREA_2 (trusted) — Krea v2 with a built-in style LoRA: fast, photoreal, uncensored. Only the prompt, ratio (the 8 standard presets) and tier reach it — 1K = 1 credit, 2K (default) = 2 credits.
  • Z_IMAGE_SPICY (trusted) — Z-Image Spicy: uncensored photoreal images from text. Only the prompt, ratio (the 8 standard presets) and mode reach it, and the provider always expands the prompt — fast (default) = 1 credit, quality = 2 credits with finer detail.
  • NANO_BANANA — Among the best for realism, and reliably renders legible in-image text (posters, signage, captions); strong real-world knowledge. Heavily censored — won't produce even mildly suggestive content. Weaker facial likeness.
  • NANO_BANANA_2_1 — Google Nano Banana 2.1: precise prompt following, consistent characters and legible in-image text (posters, infographics), up to 4K. The only model with the long formats 4:1, 1:4, 8:1 and 1:8, thinking_level and web_search. tier 1K / 2K (default) / 4K = 1 / 2 / 4 credits; thinking_level high adds 1. Censored.
  • GPT_IMAGE_2_5_FLARE / GPT_IMAGE_2_5_SUNBURST — OpenAI GPT Image 2.5: exact prompt following, legible in-image text (posters, UI mocks, infographics) and transparent cut-outs. Flare is faster, Sunburst more precise; the price is the same. The 8 standard ratios, tier 1K / 2K (default) / 4K (4K is beta), quality low / medium (default) / high and transparent_background (GPT only). Per image, low / medium / high: 1K = 1 / 1 / 2, 2K = 1 / 1 / 3, 4K = 1 / 2 / 5 credits. Censored.
  • IDEOGRAM_4_5 — Ideogram 4.5: the best lettering of our models — readable quoted text, typography and logos on posters, signs and covers. The 8 standard ratios, tier 1K / 2K (default) (no 4K), quality low / medium (default) / high and rewrite_prompt (Ideogram's Magic Prompt; it may uppercase quoted text). No negative_prompt — phrase everything positively. Per image, low / medium / high = 1 / 2 / 6 credits at either tier. Censored.
  • FLUX_KLEIN_NSFW (trusted) — The most advanced local NSFW model: produces explicit content and also works with references — bring a character's face and create an action. Slightly lower quality, occasional artifacts.

by_prompt generates exactly one image per input. For N variants, pass N input objects in the inputs array of one task (do not raise batch_size).

image_editor — The main, most flexible image tool#

Edit and composite existing images by prompt. Bring references, edit and combine them; bring your character and dress them from a reference photo. It offers both SFW and NSFW models, and LoRA presets that extend NSFW capability. Use it to keep a product or object exactly (fabric, pattern, shape) while changing the scene. This is the most capable image tool — and the default for all reference-based generation (use it instead of by_ref).

Models:

  • GENERAL_NSFW (default, trusted) — Universal default, uncensored, good facial likeness; general NSFW edits such as outfit or pose changes. Older model — occasional limb artifacts.
  • NANO_BANANA — High realism and the most precise prompt-driven edits; good at edits involving in-image text. Heavily censored (no NSFW); weaker likeness.
  • NANO_BANANA_2_1 — Google Nano Banana 2.1: precise edits and composites from up to 14 references (the first is the base image), consistent characters and legible in-image text, up to 4K. The only editor with the long formats 4:1, 1:4, 8:1 and 1:8, thinking_level and web_search; width/height are ignored. Per image: 1K / 2K (default) / 4K = 2 / 2 / 4 credits, with thinking_level high 2 / 3 / 5. Censored.
  • GPT_IMAGE_2_5_FLARE / GPT_IMAGE_2_5_SUNBURST — OpenAI GPT Image 2.5: controlled edits that keep identity, product and layout while changing one element, exact copy, up to 16 references, masked edits via mask_asset (white = area to edit) and transparent cut-outs. Flare is faster, Sunburst the most precise in edits; the price is the same. The 8 standard ratios, tier 1K / 2K (default) / 4K (beta), quality low / medium (default) / high; width/height are ignored. Per image, low / medium / high: 1K = 2 / 2 / 3, 2K = 2 / 3 / 5, 4K = 2 / 3 / 6 credits, plus 1 credit per started 3 references beyond the first 4 (the mask is free). Censored.
  • IDEOGRAM_4_5 — Ideogram 4.5: precise edits that leave untouched areas as they were, the best at fixing text and logos. Up to 5 images (4 with mask_asset): the first is edited, the rest only guide. quality low / medium (default) / high, edit_precision high for an edit that restores every pixel the prompt did not ask to change, and mask_asset (white = area to edit; leave part of the image unpainted). The result keeps the first image's size — ratio, tier, width and height do not change it. Per image, low / medium / high = 1 / 2 / 6 credits. Censored.
  • QWEN_IMAGE / QWEN_IMAGE_PRO — Aesthetic, good likeness, few artifacts; Pro adds realism. Uncensored; transforms your NSFW references rather than creating explicit content from scratch.
  • SEEDREAM_5 — Newer than the default: better prompt understanding, stronger stylization, better likeness, fewer artifacts. Uncensored; transforms NSFW references.
  • SEEDREAM_5_PRO — The newest Seedream flagship: best prompt adherence and realism of the line, up to 10 references, all 15 aspect ratios. width/height are ignored — the size comes from tier: 1K = 2 credits, 2K (default) = 3 credits per image. Uncensored; transforms NSFW references.
  • WAN_2_7_IMAGE / WAN_2_7_IMAGE_PRO — Aesthetic bodies and composition, few artifacts, precise editing; Pro = higher quality. Uncensored; transforms NSFW references. Weaker at in-image text.
  • FLUX_KLEIN_NSFW (trusted) — Local flagship for explicit NSFW with references: it knows anatomy and accepts a face reference. Slightly lower quality, occasional artifacts.
  • FLUX_KLEIN_LORA (trusted) — LoRA presets that extend NSFW capability and style templates; pass a lora_id.

Z_IMAGE_SPICY is intentionally not offered here — it is text-only and cannot accept references. For explicit anatomy on a reference, use FLUX_KLEIN_NSFW.

by_ref — Put your character into a reference photo's scene#

Bring a reference photo (the scene) and 1-4 photos of your character (character_assets: a face, a body, or several angles of the same person); the tool recreates the reference's scene — pose, outfit, composition, light — for your character.

Explicit-request-only. For most reference-based work image_editor is more flexible (prompt-steerable edits, multiple references with roles) and is the recommended choice — use by_ref only when explicitly asked for it by name.

Models: only SEEDREAM_5_PRO is available through this MCP server; it is the default, so model can be omitted.

  • SEEDREAM_5_PRO — Two modes: scene recreation (default — the reference's scene is auto-described and regenerated for the character from character_assets, 1-4 asset ids, required; output follows the reference's aspect, or set ratio from 1:1, 1:2, 2:1, 1:3, 3:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 9:21, 21:9), or edit in place (keep_pose: true — the reference itself is edited: its pose, body, composition, outfit, lighting and background stay, while the face, hair and skin tone come from character_assets; ratio is ignored). Prompts are ignored — the scene description is automatic. Exactly one image per call (batch_size ignored); flat price by tier: 1K = 2 credits, 2K = 3 credits.

facegen — Create a face from scratch#

Generate a brand-new face from structured attributes: gender, age, origin/ethnicity, body type, eye/hair/beard color, hairstyle, beard and makeup. No reference image; returns several variants per request. Strength: full parametric control over appearance. There is no free-form prompt — required fields are gender, age, origin; the rest are optional appearance fields. Use it to mint a new persona reference (for likeness of a specific person, use faceswap or by_ref). Niche — used occasionally.

photoshoot — Photoshoot from face + body references#

Bring a photo of the face and a photo of the body (without a face) of your character; the tool runs them through prepared prompt presets and returns a batch of images. Presets are grouped by type, so you can produce a set in a given style or action. The prompt describes the scene (wardrobe, location, pose, lighting, mood) — not the subject, which the references encode. Strength: reproducible, identity-preserving results with no manual prompting — and it honors a hard aspect ratio (pass ratio together with matching width/height).

Bring an image and get the same subject from different camera angles (up to 10). Use it for social-media carousels and a "3D" / product overview of an object or character. There is no prompt — angle variation is automatic; the main dial is the number of images. Niche — used occasionally.

collaber — Two characters in one frame#

Bring two characters and an optional background/location photo; the tool combines them into a single scene (1–4 images). The prompt describes their interaction and the joint scene, not their individual identities (those come from the two references). Strength: keeps both characters' likeness — a convenient preset for collabs and duets.

faceswap — Swap a face on a photo#

Bring the photo where the face should be replaced plus a face photo, and the tool swaps the character. Image only — there is no video face swap.

Models:

  • FULL_HEAD — Replaces the entire head, not just the face — use when the target's hairstyle or head shape differs strongly from the source.
  • SEEDREAM_5_PRO — Highest-quality swap. swap_type picks face (default), head or person (the whole person from the face photo carried into the reference's scene — this model only); tier 1K = 2 credits, 2K (default) = 3 credits. The aspect always follows the reference. Slow — up to ~5 minutes.
  • MULEROUTER_FACE / MULEROUTER_HEAD — Cheapest and fastest option (1 credit flat). Output is delivered as-is from the provider, with no resolution control. _FACE swaps only the face, _HEAD replaces the whole head. No persona support.

flux_klein_lora — Flux Klein with LoRA templates#

🔒 Trusted account required — available only to trusted accounts (granted automatically after your first credit purchase).

Generate or edit images with Flux Klein driven by a LoRA style template. Inputs: image_assets (1–3 reference asset_ids), lora_id (required — the LoRA template id; browse via the templates catalog), an optional short prompt (the LoRA owns the style, so keep tweaks to scene/pose), and an optional ratio (1:1, 16:9, 9:16, 4:3, 3:4, 2:3, 3:2, 21:9). Base price 3 credits.

Standalone equivalent of image_editor with model=FLUX_KLEIN_LORA; prefer image_editor unless you specifically want the dedicated entry point.

upscaler — Increase image resolution#

Brings an image up to the resolution you choose and restores detail. Best for low-resolution sources that need to be cleaned up and improved; gains are limited on an already-sharp 4K image. There is no prompt — the only meaningful choices are the version and its controls.

Versions:

  • seedvr2_3b — Adaptive resolution-aware upscale. Takes detail (high/medium/low — how much the source is pre-shrunk before upscaling, high preserves the most native detail) and target_resolution (2k/4k/8k).
  • seedvr2_7b (default) — Premium diffusion-restoration upscale for 4K/6K output. Same detail/target_resolution controls as seedvr2_3b, but target_resolution only accepts 4k/6k (sets a target megapixel budget — 4k≈16MP, 6k≈36MP — distributed per the source image's aspect ratio, not a fixed linear pixel size); also takes noise_scale (0-1, synthetic detail added during upscaling, default 0).

Video generation tools#

videogen — Generate video#

Animate a photo into video (image→video). Cost depends on the model, duration (5–30 s — past 15 s only on the [email protected] family and seedance_2_5) and resolution (480p–1080p). Supports prompt enhancement, fixed camera, LoRAs and optional audio. videogen is the image-to-video tool — a starting frame (ref_asset) is required, except for reference runs (Seedance 2.x, [email protected]) and MiniMax runs from an end frame alone; for pure text-to-video (no starting image) use text_to_video.

Quick guide: general content (best price/quality) → Seedance; newest and longest clips (to 30 s, free native audio) → [email protected]; 1080p at 30 s, or several reference videos in one call → seedance_2_5 (trusted); NSFW / explicit content → [email protected] or [email protected] (both trusted), or [email protected] (needs is_trusted); tasteful content with complex actions → Kling (censored).

Each model declares its own duration and resolution capability — filter against the user's request before suggesting one, and price every candidate with zencreator_compare_prices (which can sweep resolutions cheapest-first).

Wan — best prompt understanding and first-frame animation:

  • [email protected] — Newest Wan line; with seedance_2_5, one of only two models here that go past 15 s: durations 5 / 6 / 8 / 10 / 12 / 15 / 20 / 25 / 30 s, 480p/720p/1080p. Native audio via generate_audio at no extra cost. Animates a first frame (and, unlike [email protected], accepts a last_frame behind it) or works from multimodal references — up to 9 ref_images / 3 ref_audios (≤15 s each), mutually exclusive with last_frame. Reference videos are accepted only by video2video, never here. ratio 21:9 is rejected. Price = duration × 2.4 / 4.8 / 8 credits per second at 480p / 720p / 1080p (5 s = 12/24/40, 30 s = 72/144/240) — so 480p is a genuinely cheap tier. A run takes 15–55 minutes: always zencreator_create_task + zencreator_wait_for_task.
  • [email protected] (trusted) — Wan 3.0 with content inspection lifted; same durations, constraints and price. Trusted-only and not available in every region.
  • [email protected] / [email protected] (the nsfw entry: trusted) — The same model on a faster queue: identical inputs and output, ~40 % quicker, ~10 % dearer (5 s = 14/27/45, 30 s = 79/159/270). Behind a release flag — if the backend answers "Unrecognised model", Prime is simply not enabled for your account: fall back to [email protected].
  • [email protected] — Previous line; top prompt understanding and first-frame animation. Continuous 2–15 s, 720p/1080p. (Does not accept a last_frame keyframe — for that use [email protected], seedance_pro, or seedance_v1_5_pro; to go beyond 15 s, chain clips via ref_asset.)
  • [email protected] (trusted) — Wan 2.7 for NSFW; the best choice when you have a first frame (or a frame with an action) to animate. Uncensored. Trusted-only — requires is_trusted.
  • [email protected] / [email protected] — Cheaper and older (flash is even cheaper and faster). Duration 5, 10, or 15 s; 720p/1080p.
  • [email protected] — Sharper motion than 2.2; duration 5 or 10 s; 480p/720p/1080p.
  • [email protected] — Frame-based duration; very flexible; uncensored NSFW base. This is the backend's fallback default when no model is passed (a factual fallback, not a recommendation — prefer Seedance for unspecified general content).
  • [email protected] — Presets with action-trained LoRAs that turn a first frame into a complex action. Includes "Blink" LoRAs: bring any photo of your character and the frame morphs into the desired NSFW action. The easiest option for beginners — no prompt needed, just a photo similar to the example's first frame.
  • [email protected] — The same preset idea at Wan 2.7 quality. Presets live in their own catalog (they are not interchangeable with [email protected]'s file-weight LoRAs): browse it with zencreator_list_wan27_loras — filter by search, content_tags, or gender_tags, whose valid values come from zencreator_get_wan27_lora_tags. Pass the preset's id as loras: { "<preset id>": {} } — exactly one id, empty-object value, no scale weights. The model cannot run without a preset; prompt is optional (the preset's own default prompt is used when omitted). The catalog is filtered by your account's permissions, so NSFW and trusted-gated presets appear only once nsfw_allowed ("Show 18+ content") / is_trusted are set.

Kling — censored, but animates a first frame well; newer versions cost more and understand prompts and complex actions better. Duration 5 or 10 s, 1080p:

Seedance — uncensored; best price/quality balance for content:

  • seedance_pro_fast — Faster and cheaper, less "smart". Any integer duration 2–12 s.
  • seedance_pro — Pricier and smarter. Any integer duration 2–12 s. Supports a start+end keyframe (last_frame).
  • seedance_v1_5_pro — Best quality and result; joint audio+video, micro-expressions, first+last frame. Continuous 4–12 s.
  • seedance_2_0 / seedance_2_0_fast (_fast cheaper) — Classic first(+last) frame animation, OR native multimodal references (up to 9 images / 3 videos / 3 audios via ref_images/ref_videos/ref_audios, mutually exclusive with last_frame). Durations 5/6/8/10/12/15 s; 480p/720p only (1080p rejected). Token-based price (≈1.35/3.02 credits/s at 480p/720p, ≈1.08/2.42 _fast) — use zencreator_estimate_price rather than eyeballing it.
  • seedance_2_5 (trusted) — Newest Seedance, and the only model here that combines 1080p with 30 s. Durations 5 / 10 / 15 / 20 / 30 s only — the five the product sells of the model's fourteen; 6/8/12, 25 and the other values past 15 s are refused before the charge. 480p/720p/1080p. Same two input shapes as 2.0 but far wider: up to 30 ref_images / 10 ref_videos / 10 ref_audios (30 s per file, 30 s per kind in total), and it is the only videogen model that runs from audio alone. Reference media stays mutually exclusive with last_frame, and every reference file runs 2–30 s (video and audio alike), images 300–6000 px per side at an aspect of 0.4–2.5. With references, the optional task_type says what to do with them: reference (drive a new clip), edit (rewrite the source in place) or extend (continue it, with duration as the amount to add); the last two need exactly one ref_videos entry and run at ratio: adaptive whatever you pass. An edit returns a clip as long as its source, which is spelled duration: -1 — send that value; one of the five durations is accepted and replaced, but -1 anywhere else (another task type, another model, or text_to_video) is refused. negative_prompt is not supported. Token-based price ≈2.6 / 5.8 / 19.8 credits per second at 480p / 720p / 1080p (5 s = 13/29/100, 30 s = 78/174/595) — 1080p × 30 s is the priciest single run in the whole catalog, so quote it with zencreator_estimate_price and confirm before submitting. The model is trusted-only in full, censored route included: an untrusted account gets a 403 from the price call and from zencreator_create_task, before the task exists and before any debit, so check zencreator_get_me first. Its release flag has been open to every account since 2026-09-21. Runs long, on a 9-slot queue with a 90-minute hard timeout — always zencreator_create_task + zencreator_wait_for_task.

Grok:

  • [email protected] — Censored; animates a first frame, top image-to-video, always emits native audio. Continuous 1–10 s.

MiniMax:

  • [email protected] — MiniMax H3 Max. Censored (the provider filter is lifted for trusted accounts; a refused prompt fails and refunds), always-on native audio, durations 5 / 6 / 8 / 10 / 12 / 15 s, 480p/720p/1080p. Animates a first frame, a last_frame or both — an end frame alone needs no ref_asset — or takes references: up to 9 ref_images (ref_asset included) / 3 ref_videos / 3 ref_audios, 12 files in total, never audio alone. Price = duration × 1.25 / 2 / 4 credits per second at 480p / 720p / 1080p (5 s = 7/10/20, 15 s = 19/30/60); reference videos add their seconds.
  • [email protected] — The same at half the price (5 s = 4/5/10, 15 s = 10/15/30), frames only (an end frame alone included) — no reference mode.

Native audio: models that support it accept generate_audio: true; [email protected] always emits audio. Check each model's capabilities via zencreator_get_tool_schema.

text_to_video — Text-to-video#

🔒 Trusted account required — available only to trusted accounts (granted automatically after your first credit purchase), regardless of whether the request is SFW or NSFW.

Generates a first frame (on Flux Klein, Wan 2.2 or Wan 2.7) and then animates it — except the [email protected], seedance_2_0, seedance_2_5 and [email protected] families, which generate the clip directly from the prompt in one pass. Use this tool when there is no starting image; if you already have a frame, use videogen.

Durations are model-specific — an out-of-set duration bills a 1-credit no-op, so match them exactly:

  • [email protected] (default) — Top quality. Durations 5 / 10 / 15 s (not 8). 720p or 1080p (price ≈ 2.6 credits/s at 720p, ≈ 3.4 credits/s at 1080p — so 5 s = 13/17, 10 s = 26/34, 15 s = 39/51).
  • [email protected] — Budget. Durations 5 / 8 s (not 10/15). Resolution is ignored — flat 10 credits (5 s) / 13 credits (8 s).
  • [email protected] / [email protected] (the nsfw entry: trusted, region-limited) — Newest Wan; with seedance_2_5, one of only two ways past 15 s here, and the only one that accepts 25 s: durations 5 / 6 / 8 / 10 / 12 / 15 / 20 / 25 / 30 s, 480p/720p/1080p, free native audio via generate_audio, one-pass generation. Price = duration × 2.4 / 4.8 / 8 credits per second at 480p / 720p / 1080p (5 s = 12/24/40, 30 s = 72/144/240) — no first-frame surcharge, unlike [email protected] on this tool. ratio 21:9 is rejected. Runs 15–55 minutes.
  • [email protected] / [email protected] — The same pair on a faster queue (~40 % quicker, ~10 % dearer: 5 s = 14/27/45, 30 s = 79/159/270). Release-flagged — an "Unrecognised model" answer means it is not enabled for your account; fall back to [email protected].
  • seedance_2_0 / seedance_2_0_fast (_fast cheaper) — Generates the video directly from the prompt in one pass (no separate first-frame step). Durations 5 / 6 / 8 / 10 / 12 / 15 s. 480p or 720p only (1080p rejected). Token-based price (≈1.35/3.02 credits/s at 480p/720p, ≈1.08/2.42 _fast) — use zencreator_estimate_price rather than eyeballing it.
  • seedance_2_5 (trusted) — Newest Seedance, one-pass like 2.0 and the only model here that combines 1080p with 30 s. Durations 5 / 10 / 15 / 20 / 30 s only (6/8/12, 25 and the other values past 15 s are refused); 480p/720p/1080p. Token-based price ≈2.6 / 5.8 / 19.8 credits per second (5 s = 13/29/100, 30 s = 78/174/595) — confirm the 1080p numbers with the user before submitting. Trusted-only in full — an untrusted account is refused with a 403 before any debit; the release flag is open to every account. task_type does not apply here — it describes references, and this tool has none. Runs long: zencreator_create_task + zencreator_wait_for_task.
  • [email protected] / [email protected] (-turbo half price) — MiniMax H3 Max, one-pass, censored, always-on native audio. Durations 5 / 6 / 8 / 10 / 12 / 15 s, 480p/720p/1080p. Price = duration × 1.25 / 2 / 4 credits per second (turbo: half) — 10 s = 13/20/40, turbo 7/10/20. ratio: adaptive means 16:9 here.

video2video — Replace a character in a video#

Transfer motion / replace a character in a video using a reference video or an Instagram/TikTok URL (passed directly — no upload needed); the original soundtrack can be kept. SFW and NSFW variants. resolution (480p / 720p / 1080p) is required.

Modes:

  • kling_2_6_sfw — Handles character replacement best (censored); billed per second.
  • replace_sfw / replace_nsfw — Same character-replacement logic; replace_nsfw is uncensored.
  • animate_sfw / animate_nsfw — Motion transfer / animation; animate_nsfw is uncensored.
  • dreamactor_m2 (trusted) — Same character-replacement logic, uncensored. Trusted-only — requires is_trusted.
  • seedance_2_0_sfw / seedance_2_0_nsfw / seedance_2_0_fast_sfw / seedance_2_0_fast_nsfw (_fast cheaper) — ByteDance Seedance 2.0. Different input contract from every mode above: no Instagram/TikTok URL and no generic video_asset — the source video must be uploaded and passed via ref_videos (first entry is the main video, required), optionally alongside ref_images/image_asset/ref_audios for extra multimodal references (up to 9/3/3). prompt is required (unlike the other modes). resolution is 480p or 720p only — 1080p is rejected for these four modes.
  • seedance_2_5_sfw / seedance_2_5_nsfw (both trusted; the nsfw one is also region-limited) — Newest Seedance, one mode per policy and no _fast tier. Same asset-only contract as the 2.0 modes (source video in ref_videos[0], no Instagram/TikTok URL, prompt required), with much more room: 1080p is allowed, durations run to 30 s (the model's full list — 5/6/8/10/12/15 plus every even value to 30, not 25 — where videogen and text_to_video sell only 5/10/15/20/30), and the reference budget is 30 images / 10 videos / 10 audios. The ceiling is on the sum: duration plus the measured length of the source, rounded up to whole seconds, must fit in 30 s — so a 22 s source leaves 8 s, and a 5 s source leaves 25 s, of which the longest selectable output is 24 s, since 25 is not one of this model's values. task_type picks the behaviour — reference (a new clip driven by the source), edit (rewrite it in place; the result is as long as the source, sent as duration: -1, and is billed for twice the source's length regardless of duration) or extend (continue it, with duration as the amount to add); the last two take exactly one reference video and force ratio to adaptive. Pass task_type to the price call as well, or an edit is quoted as "source + chosen duration" instead. Price = 1.54 / 3.46 / 11.79 credits per second at 480p / 720p / 1080p over duration plus the source's seconds, with at least 9 seconds billed and never less than the same output would cost with no source at all (a 5 s source with a 5 s result = 16/35/118 credits; a 25 s result = 65/145/496). The release flag is open to every account; an untrusted one is refused with a 403 before any debit.
  • minimax_3_0 — MiniMax H3 Max, censored (the provider filter is lifted for trusted accounts; a refused run fails and refunds), open to every account. The source video must be uploaded and passed as video_asset or the first ref_videos entry — no Instagram/TikTok URL — and prompt is required. Extra references: up to 9 images (image_asset + ref_images), 3 videos including the source (≤15 s together) and 3 audios (≤15 s together), 12 files in total. Durations 5 / 6 / 8 / 10 / 12 / 15 s — always pass duration, or a video_asset run is priced as if the output were as long as the source. 480p/720p/1080p. The audio is always generated; the original soundtrack cannot be kept. Price = (duration + the reference videos' seconds, at most 15) × 1.25 / 2 / 4 credits per second at 480p / 720p / 1080p, plus 0.5 per reference image (a 5 s source with a 5 s result = 13/20/40 credits).

The uncensored modes trade some quality and prompt understanding — you may need to change the input (source video or character) to get a good result on the first try. Modes marked (trusted) are available only to trusted accounts.

lipsync — Talking head#

Bring an audio file and a first frame, and get a video in which the character speaks your audio. Up to 35 seconds; JPG/PNG under 5 MB. Use it to voice a character or avatar. Niche — used occasionally.

Models:

  • GENERAL_NSFW (default, trusted) — Specialized lipsync pipeline. Note: this GENERAL_NSFW refers to a different underlying model than GENERAL_NSFW under by_prompt / image_editor; the shared name is a backend artifact. lipsync takes audio + a first-frame portrait image (no video source) and has no text prompt.

video_upscaler — Upscale video#

Bring a medium-quality video and get a sharper, higher-resolution result via BytePlus vCube. Use it as a final polish or to restore low-quality footage. There is no prompt. Pick a quality tier (mode: standard — fast and affordable, default; or pro — premium model, ~10x the price) and a target output resolution (resolution: 1080p default, 2k, or 4k). Output is always locked to 30fps. Priced per output second by tier x resolution x duration — check zencreator_estimate_price before submitting.

video_merger — Concatenate clips#

Stitch 2–5 video clips into a single video, with a transition between each. No prompt, no model selection — it is the final assembly step after generating individual clips with videogen / text_to_video. Inputs: clips (2–5 items, each an uploaded video asset_id plus its source_duration_sec, with optional trim_start_sec / trim_end_sec to cut the clip), transition (cut / dissolve / fade / slide; default cut), keep_audio (default true), fps (24 or 30; default 30), and width / height (default 1280×720). Base price 1 credit.

See also#

  • Generation MCP tools — zencreator_create_task, zencreator_run_and_wait, zencreator_get_tool_schema, zencreator_estimate_price, zencreator_compare_prices, and the rest of the task lifecycle.
  • Concepts — tasks, calls, assets, the tool-vs-model split.
  • Workflows — end-to-end recipes (image this turn, video, NSFW preflight).