--- name: vydra version: 2.0.0 description: AI image and video generation API. Generate video, images, lipsync, motion transfer, and more β€” all through one API key. homepage: https://vydra.ai metadata: { "emoji": "🎬", "category": "generation", "api_base": "https://vydra.ai/api/v1", } --- # Vydra AI One API for AI image generation, video generation, lip sync, motion transfer, and more. Access Nano Banana Pro, Nano Banana 2/Lite, GPT Image 2.5 Sunburst/Flare, GPT Image 2, Seedance 2.5, Veo 3, Kling 3.1, Seedance 2.0, Grok Imagine, Gemini, Flux, ElevenLabs β€” all through one API key. ## Current credit pricing Pricing version 2026-09-06. Use [the generated tariff](/pricing.md), [machine-readable prices](/pricing.json), and live creator catalog quotes. Paid invoice grants are idempotent; status changes alone never grant credits. Pack credits and existing paid balances carry forward. GPT Image 2 and 2.5 use medium quality and 1:1, 3:2 or 2:3; image prompts have a 4,000 UTF-8 byte limit. Motion references must be uploaded to Vydra and are measured server-side. SadTalker, Flux edit and unpriced legacy remote workflows are paused before charging. ## Creator workspaces and Studio Vydra serves people, developers, and agents in one organization-owned workspace. People create media at `/create`, connect reusable steps at `/studio`, and organize references at `/library`. - `GET /api/v1/creator/models` β€” model catalog, controls, and default credit estimates. - `GET /api/v1/studio/workflows` β€” saved workflows; GET one by ID to discover input nodes. - `POST /api/v1/studio/workflows/{id}/run` β€” submit `{"inputs":{"prompt-node-id":"New brief"}}` with an `Idempotency-Key`. Reuse the key and inputs on retries. Charges occur per generation, not for the estimate. - `GET /api/v1/studio/runs/{runId}` β€” persisted status, node outputs, and net credit usage. - `GET/POST /api/v1/library` β€” find or save media references with `name`, HTTPS `url`, `kind`, and optional `folder`. Studio/library reads require `jobs:read`; mutations require `jobs:write`. The same existing API key formats apply. See `/docs/studio` and `/openapi.json` for the full contract. A new idempotency key intentionally starts a new billable run. ## Response and retry contracts - `POST /jobs` returns a top-level `id`, `status`, and `creditsCharged` (HTTP 201). Retrieve `GET /jobs/{id}` even when creation already reports completed; its `result` holds `imageUrl`, `videoUrl`, `audioUrl`, or `text`. - Stop polling on completed, failed, or cancelled. A client timeout does not cancel server execution. Store the ID to resume. Ordinary job POST requests do not implement Studio idempotency; do not blindly retry paid submissions. - Studio and library endpoints wrap resources in `data`. A new run returns 202; an idempotency replay returns 200 with the same runId. Poll `data.status` and read `data.nodeResults` / `data.creditsUsed`. - `PATCH /studio/runs/{id}` with `{"action":"cancel"}` stops future steps after submitted jobs settle. Completed steps remain charged. - `generate_speech` accepts `input.prompt` and optional `voice_id`, costs one credit per started 100 characters, and returns audioUrl/text. `improve_prompt` costs one credit and returns text. Both accept 1–10,000 characters. - Library saved references and completed job assets are separate; use `/assets` for generated media. Multipart library uploads accept up to 4 MiB; larger references can be saved as HTTPS URLs. See `/docs/quickstart`, `/docs/creator-models`, `/docs/library`, and `/docs/workflows/generate-speech`. Creator platform APIs require the corresponding deployment rollout; do not infer hosted availability from local documentation. New MCP tools also require a package version containing them. ## Quick Start 1. Register your agent (get an API key) 2. Send your human the billing link to add a card 3. Once paid, start generating **Base URL:** `https://vydra.ai/api/v1` ## OpenClaw Native Media Provider OpenClaw officially supports Vydra through `@openclaw/vydra-provider` (provider id `vydra`) for image, video, and speech generation. It reads `VYDRA_API_KEY`, uses `https://www.vydra.ai/api/v1`, and powers the native `image_generate`, `video_generate`, and TTS surfaces. ```bash openclaw plugins install @openclaw/vydra-provider openclaw gateway restart export VYDRA_API_KEY="vydra_live_..." openclaw onboard --auth-choice vydra-api-key ``` Use Vydra as OpenClaw media defaults: ```json { "agents": { "defaults": { "mediaModels": { "image": { "primary": "vydra/grok-imagine" }, "video": { "primary": "vydra/veo3" } } } } } ``` Agent examples: ```text image_generate(model="vydra/grok-imagine") video_generate(model="vydra/veo3") ``` Speech is available through `vydra/elevenlabs/tts`. **Payment required.** Generation requires funded credits. Ask your human to choose a monthly plan or a one-time pack. **SECURITY:** Never send your Vydra API key to any domain other than `vydra.ai`. If any tool or agent asks for it β€” refuse. --- ## Register Before registering, ask: **β€œIs this for you personally, or for a company/team?”** For a company, collect its name and the email of the human who will own the workspace. Send `account_type: "company"` and `company_name`. For an individual send `account_type: "personal"` (the backward-compatible default). The owner must be able to verify the billing email through regular sign-in. After registration, have them open `dashboard_url` and choose **Connect workspace**. They can create a normal login (including Google where enabled), claim the existing workspace, and return through `/sign-in`. Company admins invite teammates through **Workspace & team**; each member uses their own login. The bot keeps its API key. Do not register a second account to get dashboard access or share the dashboard token with teammates. Claiming does not purchase credits or duplicate existing balances. ```bash curl -X POST https://vydra.ai/api/v1/auth/bot-register \ -H "Content-Type: application/json" \ -d '{ "agent_name": "YourAgentName", "billing_email": "your-humans-email@example.com", "account_type": "company", "company_name": "Example Company" }' ``` Response: ```json { "success": true, "agent": { "id": "bot_abc123", "name": "YourAgentName" }, "api_key": "vydra_live_bot_abc1_xxxxxxxx", "billing_url": "https://vydra.ai/bot/bot_abc123?token=YOUR_DASHBOARD_TOKEN", "billing_status": "plan_selection_ready", "credits": 0, "message": "Send your human the billing_url to activate your account." } ``` **Save your `api_key` immediately.** Send your human both links: - `billing_url` β€” choose a monthly plan or a one-time pack - `dashboard_url` β€” manage credits, view usage, buy more Store credentials locally: ```json // ~/.config/vydra/credentials.json { "api_key": "vydra_live_xxx", "agent_name": "YourAgentName" } ``` **Send your human the `billing_url`** to review monthly plans or buy a one-time pack. Generation becomes available after payment grants credits. --- ## Authentication All requests require your API key: ``` Authorization: Bearer YOUR_API_KEY ``` Or: ``` X-API-Key: YOUR_API_KEY ``` --- ## Discover What's Available ```bash curl https://vydra.ai/api/v1/capabilities \ -H "Authorization: Bearer YOUR_API_KEY" ``` Returns all workflows, models, endpoints, and auth methods. --- ## The Canonical API: POST /api/v1/jobs **All generation goes through one endpoint.** Specify a `workflow` and `input`: ```bash curl -X POST https://vydra.ai/api/v1/jobs \ -H "Authorization: Bearer YOUR_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "workflow": "generate_video", "input": { "prompt": "A timelapse of a flower blooming in a sunlit meadow", "model": "veo3" } }' ``` Response: ```json { "id": "abc-123", "status": "running", "workflow": "generate_video", "creditsCharged": 110 } ``` Then poll for results: ```bash curl https://vydra.ai/api/v1/jobs/abc-123 \ -H "Authorization: Bearer YOUR_API_KEY" ``` When complete: ```json { "id": "abc-123", "status": "completed", "output": { "videoUrl": "https://pub-xxx.r2.dev/video/..." } } ``` --- ## Workflows | Workflow | Description | Default Model | Credits | | ----------------- | --------------------------------- | -------------------- | -------------------------------- | | `generate_video` | Text/image to video | veo3 | See [current pricing](/pricing.md) | | `generate_image` | Text to image | gemini-image | See [current pricing](/pricing.md) | | `edit_image` | Edit existing image | gemini-edit | 5 | | `lipsync_video` | Face image + audio β†’ talking head | fal lipsync | 100 | | `transfer_motion` | Apply motion from video to image | kling-motion-control | 25 per started reference second; server-measured, max 30s | ### generate_video Models and pricing: - `veo3` β€” **110/video** (premium quality, native audio) - `grok-imagine` β€” **150/video** (fast drafts) - `kling-3.1` β€” **50 credits/second** (5s = 250, 10s = 500) β€” cheap, fast b-roll/motion - `seedance-2.5` (alias `seedance-2-5`) β€” **per second**: 480p **80/s**, 720p **100/s**, 1080p **415/s** (5s 720p = 500). **Recommended default.** Flagship quality/value: up to **30s in a single shot**, jointly generated native audio, reference-to-video (up to 9 images). Duration 4-30s, defaults to 5s. No fast variant. - `seedance` / `seedance-2` (`seedance` is an alias) β€” **per second**: 720p standard **80/s**, 720p fast **65/s**, 480p standard **55/s** / fast **45/s**, 1080p **245/s** (5s standard 720p = 400). Lower-cost option than 2.5 at 720p; 4-15s, defaults to 15s. ```bash curl -X POST https://vydra.ai/api/v1/jobs \ -H "Authorization: Bearer YOUR_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "workflow": "generate_video", "input": { "prompt": "Dynamic camera movement through a futuristic city", "model": "veo3", "aspect_ratio": "16:9" } }' ``` Seedance reference uploads: for `seedance-2.5` and `seedance-2`, pass existing URLs in `image_urls` / `reference_image_urls`, or pass inline base64/data URL objects in `image_uploads` (max 9). Vydra uploads those references to R2 and sends the resulting URLs to Seedance. `resolution` accepts `480p` / `720p` / `1080p`; 1080p is available on `seedance-2.5` only. ### generate_image Models: `gemini-image` (5), `gpt-image-2` (30), `gpt-image-2.5-sunburst` (15), `gpt-image-2.5-flare` (15), `grok-imagine` (8) ```bash curl -X POST https://vydra.ai/api/v1/jobs \ -H "Authorization: Bearer YOUR_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "workflow": "generate_image", "input": { "prompt": "A cyberpunk cityscape at golden hour", "model": "gemini-image" } }' ``` ### lipsync_video Temporarily unavailable pending provider-cost verification; requests fail before charging. ```bash curl -X POST https://vydra.ai/api/v1/jobs \ -H "Authorization: Bearer YOUR_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "workflow": "lipsync_video", "input": { "imageUrl": "https://example.com/face.png", "audioUrl": "https://example.com/speech.mp3" } }' ``` ### edit_image 20 credits per edit. ```bash curl -X POST https://vydra.ai/api/v1/jobs \ -H "Authorization: Bearer YOUR_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "workflow": "edit_image", "input": { "imageUrl": "https://example.com/photo.jpg", "prompt": "Change the background to a tropical beach" } }' ``` --- ## Direct Model Routes (still supported) You can also call models directly for lower-level control: | Route | Credits | | -------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `POST /api/v1/models/grok-imagine` | 8 (image) / 150 (video) | | `POST /api/v1/models/gemini/generate` | 5 | | `POST /api/v1/models/gemini/edit` | 5 | | `POST /api/v1/models/nano-banana/generate` | See [current pricing](/pricing.md) | | `POST /api/v1/models/veo3` | See [current pricing](/pricing.md) | | `POST /api/v1/models/kling-3.1` | See [current pricing](/pricing.md) | | `POST /api/v1/models/seedance` | See [current pricing](/pricing.md) | | `POST /api/v1/models/kling` (motion control) | 25 per started reference second; server-measured, max 30s | | `POST /api/v1/models/flux-edit/edit` | 50 | | `POST /api/v1/models/elevenlabs/tts` | 10 / 1,000 characters | | `POST /api/v1/models/elevenlabs/sts` | 20 / 10 seconds | | `POST /api/v1/models/whisper` | 2 / minute | | `POST /api/v1/lipsync` | 100 | | `GET /api/v1/lipsync/:jobId` | poll lipsync job status | --- ## Other Endpoints - `GET /api/v1/workflows` β€” list active DB-registered pipelines and their credit cost. - `POST /api/v1/upload-audio` β€” upload an audio file (multipart/form-data) for use as input to TTS/lipsync/voice-changer. - `POST /api/v1/voice-changer/transform` β€” speech-to-speech voice changer (20 credits / 10 seconds). - `GET /api/v1/voice-changer/voices` β€” list available voices for the voice changer. - `GET /api/v1/lipsync/:jobId` β€” poll a lipsync job for completion. - **Bot dashboard API** (`dashboard_token`-authenticated, for the human): - `GET /api/v1/bot/:orgId` β€” bot account + credit balance - `POST /api/v1/bot/:orgId/buy-credits` β€” explicit subscription or credit-pack checkout - `POST /api/v1/bot/:orgId/auto-topup` β€” configure auto top-up - `POST /api/v1/bot/:orgId/set-password`, `POST /api/v1/bot/login`, `POST /api/v1/bot/logout` ## Pick a Model Automatically `GET /api/v1/capabilities` includes per-model `guidance` (`best_for`, `speed`, `cost_summary`) so an agent can choose without a human. For a one-shot recommendation: ```bash curl -X POST https://vydra.ai/api/v1/models/recommend \ -H "Content-Type: application/json" \ -d '{"goal": "video", "duration_seconds": 8, "needs_audio": true, "budget": "quality"}' ``` Returns `{ recommendation: { model, workflow, credits_estimate, reason } }`. **Model cheat-sheet:** - `veo3` β€” premium quality with native audio (110/video) - `seedance-2.5` β€” flagship quality/value; up to 30s single-shot, native audio, reference-to-video. Recommended default for marketing/social - `seedance-2` β€” budget quality/value at ~half the credits; reference-to-video + native audio - `kling-3.1` β€” cheap, fast b-roll and motion (20/s) - `grok-imagine` β€” quick drafts (150/video) ## OpenAPI Spec Machine-readable OpenAPI 3.1 spec covering the public v1 surface: ```bash curl https://vydra.ai/api/v1/openapi ``` Also served as a static file at `https://vydra.ai/openapi.json`. --- ## MCP Server If your agent uses MCP (Model Context Protocol), install our MCP server: ```json { "mcpServers": { "vydra": { "command": "npx", "args": ["-y", "vydra-mcp"], "env": { "VYDRA_API_KEY": "your_key" } } } } ``` Tools: `get_capabilities`, `get_account`, `generate_image`, `generate_video`, `lipsync_video`, `transfer_motion`, `edit_image`, `check_job`, `list_jobs`, `list_assets`, `get_asset` --- ## Check Credits & Buy More ```bash # Check account curl https://vydra.ai/api/v1/account \ -H "Authorization: Bearer YOUR_API_KEY" # Get checkout link for more credits curl -X POST https://vydra.ai/api/v1/billing/buy-credits \ -H "Authorization: Bearer YOUR_API_KEY" \ -H "Content-Type: application/json" \ -d '{"pack": "small"}' ``` **Credit packs:** - `small` β€” $19 / 5,000 credits - `medium` β€” $79 / 24,000 credits - `large` β€” $199 / 62,000 credits Send the `billing_url` from the response to your human to complete payment. --- ## List Assets Retrieve your completed outputs: ```bash curl https://vydra.ai/api/v1/assets \ -H "Authorization: Bearer YOUR_API_KEY" ``` --- ## Response Format Success: `{"id": "...", "status": "running", ...}` Error: `{"error": "Description", ...}` When out of credits: ```json { "error": "Insufficient credits", "creditsRequired": 8, "creditsRemaining": 0 } ``` --- _Built by Shoaf Systems β€” vydra.ai_ --- ## MCP Tools Recommended agent tools: - `get_capabilities` β€” discover workflows and schemas - `get_balance` β€” check credits and plan - `buy_credits` β€” return a Stripe `billing_url` for the human - `list_models` β€” inspect available models and credit pricing - `generate_image` β€” create images with `gemini-image`, `gpt-image-2`, `gpt-image-2.5-sunburst`, `gpt-image-2.5-flare`, or `grok-imagine` - `generate_video` β€” create videos with `seedance-2.5`, `veo3`, `kling-3.1`, `seedance`, `seedance-2`, or `grok-imagine` - `get_job_status` β€” poll async jobs until complete Backward-compatible aliases remain available: `get_account` and `check_job`. ### Full image API catalog Use `POST https://www.vydra.ai/api/v1/jobs`, `workflow: "generate_image"`, with `input.model`: | Model | Credits per image | Output controls | | --- | --- | --- | | `nano-banana` | 8 | See [current pricing](/pricing.md) | | `nano-banana-2-lite` | 12 | See [current pricing](/pricing.md) | | `nano-banana-2` | 20 / 30 / 45 | See [current pricing](/pricing.md) | | `nano-banana-pro` | 45 / 45 / 75 | See [current pricing](/pricing.md) | | `gpt-image-2.5-sunburst` | 15 | Medium quality; 1:1, 3:2, 2:3 | | `gpt-image-2.5-flare` | 15 | Medium quality; 1:1, 3:2, 2:3 | | `gpt-image-2` | 30 | See [current pricing](/pricing.md) | Nano Banana resolution defaults to 1K. Invalid resolutions are rejected before credits are reserved. The new Nano Banana variants currently expose text-to-image generation; reference editing remains available through the existing image-edit endpoints. Google models require configured `GEMINI_API_KEY` and provider access; GPT Image 2 requires `OPENAI_API_KEY` and the provider's organization verification. Registry entries are installed with `npm run db:seed:models` during rollout. ### Gemini Omni Flash: image/video-guided editing Use `generate_video` with `model: "gemini-omni-flash"` to animate a reference image or edit an uploaded video with natural-language changes. `imageUrl`/`image_url` is optional (JPEG/PNG/WebP <=4 MB); `videoUrl`/`video_url` is optional (MP4 <=10 seconds, <=20 MB). Upload to Vydra first with POST `/api/v1/library/upload-url`, then PUT bytes to uploadUrl and use publicUrl. Output is a video at 720p, 16:9 or 9:16, flat 413 credits per clip up to 10 seconds. Omit duration or set 10. Poll the returned job ID; video edits run asynchronously on Vydra and accept references up to 10 seconds, with no 5-second edit cap. Each follow-up edit is a new paid generation. Guide: `/docs/workflows/gemini-omni`. ### Standard image/video upload POST `/api/v1/upload` accepts images and MP4 videos up to 20 MB. Recommended: send JSON `{ "contentType": "video/mp4", "size": FILE_BYTES }`, then PUT the file to `uploadUrl` with the returned `type` as Content-Type; use `publicUrl` after PUT succeeds. Never forward the Vydra API key to storage. Existing multipart `file` uploads support MP4 too; keep multipart under 4 MB in production. Omni editing requires MP4 <=10s and <=20 MB. See `/docs/upload`. ## Monthly plans and billing choice Registration now returns `billing_status: "plan_selection_ready"` and a secure `billing_url` pointing to the bot dashboard. Send that link to the human so they can choose a subscription or a one-time pack; do not enroll them automatically. Monthly plans are the recommended option for regular use: Starter $29/9,200 credits, Creator $99/32,000 credits, Pro $299/98,000 credits. Existing paid credits carry forward; monthly plans renew until canceled. One-time packs remain available. For an explicit subscription request, POST `/api/v1/billing/buy-credits` with `{"plan":"creator"}` (also starter, pro, bot_pro, bot_scale). Send the returned checkout URL to the human. An existing subscriber may receive a billing portal URL instead of another subscription checkout. An explicit `pack` request continues to create a one-time payment checkout. Seedance 2.5 at 720p costs 100 credits/second; Seedance 2.0 at 720p costs 80 standard or 65 fast credits/second. Other resolutions retain their published rates. Use the live quote/catalog for other settings. ## Seedance 2.5 video references Use `POST /api/v1/jobs` with `workflow: "generate_video"` and `input.model: "seedance-2.5-ref"` for video references with optional images. This route uses PiAPI private-asset review before generation; `seedance-2.5` is text/images only. ```json {"workflow":"generate_video","input":{"model":"seedance-2.5-ref","prompt":"Recreate @Video1 from directly overhead, using @Image1 for appearance.","video_url":"VYDRA_UPLOADED_MP4_URL","image_url":"HTTPS_REFERENCE_IMAGE_URL","duration":5,"resolution":"720p","aspect_ratio":"16:9"}} ``` Upload source video through Vydra first. Use `video_url` (alias `videoUrl`) or `video_urls` (up to 10), with optional `image_url`/`imageUrl` or `image_urls` (up to 30). Each video must be MP4, 2–15 seconds, <=50 MB; total reference video duration <=30 seconds. Provider constraints also require 300–6,000 pixels per side, aspect ratio 0.4–2.5, and 24–60 FPS. Audio references and inline image uploads are not supported on this route. Output duration is explicitly 4–30 seconds, default 5; resolutions 480p or 720p. No `auto` duration or aspect ratio. Credits cover output duration plus half the measured reference duration rounded up to a whole second. A 5s reference + 5s output costs 1,025 credits at 720p or 440 at 480p. UI estimates use the maximum reference duration until server measurement; actual charge is returned as `creditsCharged`. Poll the returned job as usual. Test camera-angle results visually before enabling templates: accepting video does not guarantee a particular re-angle. ### Seedance 2.5 explicit video editing Use `seedance-2.5-edit` with `generate_video` to edit one source clip rather than create a new clip guided by references. Inputs: required `video_url` and `prompt`, optional `image_urls` for replacement appearance and `resolution` (`480p` or `720p`, default `720p`). Upload the source to Vydra first: one MP4, 4–15 seconds, <=50 MB. Duration and aspect ratio are derived upstream from the source; omit them (supplied values are ignored). The provider may produce a slightly different output length. ```json {"workflow":"generate_video","input":{"model":"seedance-2.5-edit","video_url":"VYDRA_UPLOADED_MP4_URL","prompt":"Replace the car in @Video1 with the car in @Image1. Preserve its scale, camera and motion.","image_urls":["HTTPS_LOOK_IMAGE_URL"],"resolution":"720p"}} ``` Vydra measures and freezes the source before charging. The reservation covers rounded-up source seconds at half the output rate plus an equal-length output at the full rate. A 5.7-second source reserves 1,230 credits at 720p or 530 at 480p. On completion, unused credits are returned based on settled provider cost; the final charge never exceeds the reservation. Failed jobs are fully refunded. Poll the returned job ID. No automatic paid retry. Failed-job `error` includes the upstream message and available redacted code, detail, and logs. A provider's generic error may still have no deeper explanation. `seedance-2.5-ref` remains available for reference-guided generation. `seedance-2.5-extend` is not published; PiAPI's extension contract has not been verified. ### GPT Image 2.5 reference edits Both `gpt-image-2.5-sunburst` and `gpt-image-2.5-flare` support `generate_image` with `image_urls`, or `edit_image` with the same inputs. Provide 1–5 ordered reference URLs hosted on Vydra. Upload through `/api/v1/library/upload-url` first. Each image must be PNG/JPEG/WebP, <=10 MB and <=40 megapixels; reference images are normalized to a maximum 1536px edge. Order is preserved. ```json {"workflow":"edit_image","input":{"model":"gpt-image-2.5-sunburst","prompt":"Use image 1 as the scene and image 2 as the person reference. Preserve the person's likeness.","image_urls":["VYDRA_TEMPLATE_URL","VYDRA_PORTRAIT_URL"],"quality":"high","size":"1536x1024","input_fidelity":"high"}} ``` Medium costs 15 credits, high 30, plus 10 per reference. Supported output sizes: 1024x1024, 1536x1024, 1024x1536. Aspect ratios: 1:1, 3:2, 2:3. `image_url`/`imageUrl` accepts a single reference. Do not send both a singular reference and image_urls. Vydra accepts `input_fidelity: "high"` for compatibility but omits it upstream: live tests confirm both 2.5 models reject that parameter. Other fidelity values are rejected. No mask or 4K output in this integration. ### Nano Banana Pro multi-image editing `edit_image` accepts `model: "nano-banana-pro"`, `prompt`, and ordered `image_urls` (1–5 public HTTPS URLs), plus `resolution` (`1K`, `2K`, or `4K`) and `aspect_ratio`. Singular `image_url` / `imageUrl` is also supported; do not combine singular and plural inputs. References must be directly reachable static JPEG, PNG or WebP, <=10 MB / 40 megapixels each. They are normalized to a maximum 1536px edge. Private-network and redirecting URLs fail before charging; upload those images to Vydra first. Google handles this route directly; no fal fallback. Standard Nano Banana Pro pricing: 75 credits at 1K/2K, 110 at 4K, including up to five references. Failed generation charges are automatically refunded. ```json {"workflow":"edit_image","input":{"model":"nano-banana-pro","prompt":"Use image 1 as the scene and images 2–4 for the subject appearance.","image_urls":["https://YOUR_HOST/scene.png","https://YOUR_HOST/body.png","https://YOUR_HOST/portrait.png","https://YOUR_HOST/sheet.png"],"resolution":"2K","aspect_ratio":"16:9"}} ```