Blog
AI Image & Video API Guides
Hands-on model tests, production tutorials, prompt techniques, and current API pricing for generative media teams.
Seedance 2.0 API Pricing: Mini, Fast and Standard per Second
Mini 720p is $0.0586/s, Fast $0.1772/s, standard $0.293/s; the ext route is $0.134/s at 720p, $0.342/s at 1080p and $0.694/s at 4K. A 5-second 720p clip with audio: Mini $0.293, Fast $0.886, standard $1.465. Audio costs nothing extra. Draft on Mini 480p, render finals on ext 1080p, and use standard when you need a reference video.
How to Use the kling-3.0-omni/transformation API: Video Restyle with curl and Python
Restyle an existing clip with Kling 3.0 Omni's video-to-video model through hiapi's unified async endpoint POST /v1/tasks — a working curl request, a complete Python client that polls and downloads, and the production callback pattern. The schema is strict and live-verified: prompt, exactly one public source video URL, and a resolution of 720p, 1080p, or 4k. Restyle strength is controlled by the prompt, not a numeric parameter.
How to Use gpt-6.1-sol via the hiapi API: curl, Python, and a Working Request
gpt-6.1-sol is a reasoning-grade text model you call through hiapi's POST /v1/responses endpoint — not /v1/chat/completions. Streaming is mandatory: set stream:true and read Server-Sent Events; stream:false returns a 400 error. This guide has copy-paste curl and Python examples, SSE parsing, system prompts via instructions, reasoning.effort control, and the common errors.
HiAPI 502 / 503 Upstream Errors: Causes and How to Fix Them
A 502 Bad Gateway or 503 Service Unavailable from HiAPI is a server-side 5xx response: the request was not malformed, so the fix is an idempotent retry, not a code change. 503 is HiAPI's one documented 5xx — its body carries error_code TEMPORARILY_UNAVAILABLE and the docs say to retry later with exponential backoff; 502 (and 504) are not in the documented status table and come from a gateway/proxy on the path. Distinguish both from a TASK_FAILED or TASK_TIMEOUT result, which arrive as HTTP 200 with status "fail" — the task ran and failed, rather than the request being rejected at the gateway. Send an Idempotency-Key header on POST /v1/tasks so a retry after a 502/503 returns the original taskId (Idempotent-Replay: true) instead of creating a duplicate task or charge; failed tasks are never charged.
Kling 3.0 Omni Reference-to-Video: 3 Copy-Paste Prompt Recipes With Real Clips
Kling 3.0 Omni reference-to-video keeps specific characters and products consistent: pass 1–7 reference images in image_urls, or define named elements and call them as @name. Each element needs a name without spaces, a short description and 2–4 images of the same subject. Billing is per second: 720p from $0.10/s, 1080p from $0.129/s, 4k $0.479/s; sound and reference videos cost more. The three recipes below (product, character, character + product) were generated with the exact requests shown.
FLUX.2 vs Nano Banana 2: Which hiapi Image API Should You Use?
FLUX.2 is the control-and-clean choice: tight prompt adherence, an art-directable studio look, and cheap 0.5 MP draft iteration. Nano Banana 2 is the scene-and-scale choice: richer real-world context, a 4K tier, in-model editing, and a lower per-image bill. Both rendered the same prompts faithfully, down to an exact three-bean count; pick by the job, and A/B them with one key on hiapi.
Best Stability AI Alternatives in 2026: Choosing an Image and Video Generation API
Stability AI's real edge is open weights and self-hosting the Stable Diffusion lineage (SDXL, SD 3.5); no managed API, including HiAPI, replaces running the weights on your own hardware. Its hosted Developer Platform is image-first (Stable Image Ultra/Core, SD 3.5 tiers); it is not where you add hosted text-to-video or non-SD image models. HiAPI is the broadest managed move: 60+ image, video, audio and text models under one key, OpenAI-compatible text, and a native MCP server. Replicate hosts the Stable Diffusion lineage directly and the widest open catalog; fal.ai leads on inference speed; Runware competes on lowest per-image cost. Pricing moves often: Stability uses credits (1 credit ~ $0.01), Stable Image Core ~ $0.03 and Ultra ~ $0.08; HiAPI bills per image from ~ $0.005. Confirm live numbers before budgeting.
HiAPI Image Input Errors: Pass a Public URL, Not Base64
HiAPI image-to-image and reference-image models read the source image from a public HTTP(S) URL — not base64 and not a private address. A bad image input usually passes submit (200 + taskId) and only fails on poll: INVALID_REQUEST 'could not be downloaded' for an unreachable URL, or TASK_FAILED 'task failed' for base64. The image-input field name varies by model: input_urls (gpt-image-2), image_urls (seedream-5.0-pro), image_input (Nano-Banana-Pro). Use the image-to-image route id, not the bare text-to-image id, or the image field is rejected with a 400 additional-properties error. Fix: host the image at a public, no-auth URL, put it in the model's documented field, then poll the task to read data.error.message.
HiAPI TASK_FAILED Error: Causes and How to Fix It
A terminal TASK_FAILED returns HTTP 200 — the failure lives in data.status:"fail", not in the HTTP code. "TASK_FAILED" is the data.error.code; data.status itself is always "success" or "fail". The most common cause is a broken or non-public input image/video URL; fix the URL and verify it with curl -I. Tell it apart from 400/401/402 rejections (no taskId is returned) and from timeouts (the task never reaches a terminal state). Always poll GET /v1/tasks/{taskId} and read data.error.message — it names the cause.
MiniMax H3 API Pricing: H3 and H3 Max Costs
H3 2K costs $0.1189 per second; H3 Max costs $0.072 at 480P or $0.115 at 768P. A 10-second clip costs $1.189, $0.72 or $1.15; $10 buys 16, 27 or 17 whole five-second clips, totaling 80, 135 or 85 seconds. H3 provides 2K and multimodal references; Max offers 480P, 768P, and independent first- or last-frame inputs.
Build with the models you just read about.
Compare image, video, and audio models, then use the same API workflow in production.