Changelog

July 2026 — Kling 3.0 Motion Control (Kuaishou)

Kuaishou's Kling 3.0 Motion Control lands on POST /v2/videos/generate as kling-3.0-motion-control, backed by kwaivgi/kling-v3-motion-control on Replicate and served by the existing replicate-kling provider.

Transfers motion from a reference video onto the character in a reference image — both are required in klingParams.
Adds mode (std 720p / pro 1080p), character_orientation (image caps the clip at 10s, video at 30s), and keep_original_sound to klingParams.
Per-second pricing billed on the reference video's measured duration: 12/s for std, 21/s for pro. Clips outside the allowed range are rejected before any credits are charged.

See Credit costs for details.

July 2026 — Kling 3.0 Omni (Kuaishou)

Kuaishou's unified Kling model lands on POST /v2/videos/generate as kling-3.0-omni, backed by kwaivgi/kling-v3-omni-video on Replicate and served by the existing replicate-kling provider.

One model for text-to-video, image-to-video, reference-to-video, and video editing/restyling with native audio.
Adds reference_images (up to 7; refer to them in the prompt as <<<image_1>>>, …), reference_video + video_reference_type (feature for style/camera, base for editing), and keep_original_sound to klingParams. multi_prompt is a JSON array of { prompt, duration } shots (max 6).
generate_audio is mutually exclusive with reference_video; when a reference video is set, video_reference_type is required and reference_images is capped at 4.
Same standard/pro/4K modes and per-second pricing as Kling 3.0: 29/s standard, 39/s pro, 72/s 4K (+audio: 43/s standard, 58/s pro; 4K bills 72/s with or without audio).

See Credit costs for details.

July 2026 — Stable Diffusion 3 & 3.5 (Stability AI)

Stability AI's Stable Diffusion models land on POST /v2/images/generate, backed by the stability-ai/* official models on Replicate and served by the new replicate-stable-diffusion provider.

stable-diffusion-3.5-large: highest-quality SD 3.5 tier. 12 credits per image.
stable-diffusion-3.5-large-turbo: near-Large quality in ~4 steps for fast, cheap generation. 8 credits per image.
stable-diffusion-3.5-medium: efficient mid-size SD 3.5 model. 7 credits per image.
stable-diffusion-3: the original SD 3 Medium MMDiT model with strong prompt adherence and typography. 7 credits per image.
All four share stableDiffusionParams (aspect_ratio, cfg, steps, seed, negative_prompt, image, prompt_strength, output_format, output_quality) and support image-to-image via image.

See Credit costs for details.

July 2026 — Nano Banana 2 Lite (Google)

Google's fastest, most cost-efficient Nano Banana model lands on POST /v2/images/generate as nano-banana-2-lite, backed by google/nano-banana-2-lite on Replicate and served by the same Nano Banana provider.

Built on Gemini 3.1 Flash Lite Image — text-to-image in roughly 4 seconds, ideal for prototyping and high-volume workflows.
Outputs at 1K resolution only (no resolution parameter). Supports 14 aspect ratios, jpg/png output, image input, and Google Search / Image Search grounding via nanoBananaParams.
Flat pricing: 7 credits per image. See Credit costs.

June 2026 — Happy Horse 1.0 & 1.1 (Alibaba)

Alibaba's Happy Horse video models land on POST /v2/videos/generate, served across both Replicate and Alibaba Cloud (DashScope) with automatic failover. Both bill per second of output by resolution and generate synchronized audio.

happyhorse-1.0: text-to-video and image-to-video (single first-frame image), 720p/1080p, 3–15 s. Pricing: 24/s @720p, 48/s @1080p.
happyhorse-1.1: adds reference-to-video — pass up to 9 images and reference them in the prompt as [Image 1], [Image 2], etc. Stronger motion and character consistency. Pricing: 24/s @720p, 31/s @1080p.
Mode is selected automatically: 1.0 from the presence of image; 1.1 from the number of images (0 = T2V, 1 = I2V, 2–9 = R2V).

See Credit costs for details.

April 2026 — GPT Image 2 (OpenAI)

OpenAI's new SOTA image model lands on POST /v2/images/generate as gpt-image-2, alongside the existing gpt-image-1.5. Backed by openai/gpt-image-2 on Replicate and exposed through the same replicate-gpt-image provider.

Same dual-mode contract as 1.5: omit input_images for text-to-image, pass one or more URLs to edit / compose. The model always processes inputs at high fidelity.
New gptImage2Params.moderation (auto / low) for less restrictive content filtering.
webp is the default output format.
No transparent background — keep using gpt-image-1.5 for transparent PNGs.
Pricing per output image: low 4 / medium 9 / high 23 / auto 23, multiplied by number_of_images. See Credit costs and the gpt-image-2 reference.

April 2026 — AI Photos (LoRA training)

A brand-new family of endpoints lets you train a Flux LoRA from 15-30 subject photos and then generate AI photos of that subject from any text prompt.

POST /v2/loras — multipart upload + start training. The API auto-captions every image with a vision model, packages the dataset, and runs the LoRA training. Returns 202 with the new LoRA id.
GET /v2/loras — paginated list of your team's LoRAs.
GET /v2/loras/:id — single lookup; status transitions through PENDING → UPLOADING → TRAINING → READY (or FAILED). Optional webhookUrl fires on terminal state.
DELETE /v2/loras/:id — soft-archive.
POST /v2/images/generate — new model flux-lora: pass loraId + a prompt and the API resolves the trained weights and prepends the trigger word for you. 1-4 outputs per request.

Pricing is flat: 2 credits for create (upload+caption+zip), 255 credits for training, 2 credits per image for inference. See Credit costs → AI Photos and the AI Photos overview.

April 2026 — Wan 2.7

The Wan 2.7 family ships across both Replicate and Alibaba Cloud, with a brand-new video-editing route:

POST /v2/videos/generate — wan-2.7: combined T2V + I2V at 720p/1080p, 2–15 s. New first-and-last-frame (last_frame) and clip-continuation (first_clip) anchors. Audio is auto-generated when audio is omitted.
POST /v2/videos/generate — wan-2.7-r2v: reference-to-video. Up to 5 reference_images and/or 5 reference_videos, plus shot_type for multi-shot narratives.
POST /v2/videos/edit (NEW) — wan-2.7-videoedit: prompt-driven restyle, outfit swap, and background replacement on existing clips. The route probes input duration via ffprobe when duration is omitted.

All three are billed per second of output. See Credit costs → Wan 2.7.

April 2026 — Editing endpoints

Four new endpoints for editing existing media, on dedicated routes alongside generation:

POST /v2/images/upscale — Topaz Image Upscale for realistic photos and Clarity Upscaler for creative AI art.
POST /v2/images/background-remove — Bria RMBG 2.0 for production-quality alpha mattes and 851-labs for cheap, high-volume jobs.
POST /v2/images/edit — Flux Fill Pro for masked inpainting and preset-driven outpainting (mode-switched).
POST /v2/videos/upscale — Topaz Video Upscale with up to 4K / 60 fps output and per-second pricing.

v2.0.0 — March 2026

Initial release of the Apiframe v2 API.

Image generation

Midjourney model with aspect ratio support
Nano Banana model with image-to-image and output format options

Video generation

Kling 2.6 Pro with 5s/10s duration, start image, audio generation, negative prompts

Music generation

Suno with custom lyrics mode, instrumental mode, style control, and model version selection (V4 through V5)

Platform

API key authentication with X-API-Key header
Webhook notifications for job progress, completion, and failure
Cursor-based pagination on all list endpoints
Idempotency key support for generation requests
Credit-based billing with automatic refunds on failures
Team management with role-based access

On this page