Open Cogs
Example recipes with verified defaults
Each card links to a canonical input/output showcase. Deploy the same template name with the bundled JSON — HTTP, CLI, desktop, and MCP stay in parity.
Refactor this Go handler to stream results, then explain the tradeoff in two sentences.
Prompt → reasoning, code, tool calls
Qwen3.8-27B FastServe
Qwen3.8-27B on SGLang with a C launcher that reads the GPU and picks the fastest profile for it: NVFP4 on Blackwell, FP8 on Hopper and GB300, DSPARK speculative decoding everywhere. Answers the OpenAI API and Cog predictions on one port.
masterpiece, best quality, 1girl, wind-swept cape, luminous city at dusk, detailed background, dramatic illustration
Prompt → anime illustration
Anima 2.9B
Run the expanded 2.9B Anima diffusion transformer through a native BF16 Cog with TF32, SDPA, batched guidance, a precompiled torch.compile cache, and a RunPod worker that returns to zero after the burst.
Walk forward, turn left, then break into a joyful run.
Text → live pose stream
ARDY Motion
Direct a humanoid with natural-language motion, waypoints, and WASD controls. ARDY streams structured pose frames at 20 FPS for games, avatars, and animation tools.
Portrait + audio → character video
LiveAvatar
Turn one character image and a speech or singing track into expressive long-form video with face, body, camera, and scene guidance.
Image → browser-ready 3D
Voxel 3D
Isolate an object, reconstruct its surface with TripoSR, optionally voxelize it, and return a portable GLB ready for the web or a game engine.
Image + mask → clean plate
LaMa Cleaner
Remove unwanted objects with a white mask. LaMa fills the selected region, feathers the edge, and preserves pixels outside the repair.
Image → depth-powered passes
Depth VFX
Run Depth Anything V2 once and export a depth map, heatmap, normal pass, or a subtle looping parallax move.
Raster → editable SVG
Vectorizer
Convert logos, posters, photos, and pixel art into compact SVG paths with VTracer’s deterministic multithreaded Rust core.
Small image → restored 4× image
Image Upscaler
Restore detail with Real-ESRGAN at 2× or 4×, select general or anime tuning, and optionally enhance faces with GFPGAN.
Song → clean stems
Vocal Separator
Split a track into vocals and instrumental, or export all four Demucs stems as a ZIP for remixing and post-production.
Image → transparent cutout
Background Remover
Remove a photo or illustration background to a transparent PNG with a warm, reusable segmentation session.
Prompt → accelerated video
Wan 2.2 T2V + CG-Taylor
Generate video with Wan 2.2 A14B using NF4 dual-stage transformers and confidence-gated Taylor caching for training-free acceleration.
Image + prompt → accelerated video
Wan 2.2 I2V + CG-Taylor
Animate a reference frame while preserving its aspect ratio. The I2V image uses the same R2-backed weight path and 48 GB-friendly NF4 runtime.
Character image + driving video → restyled motion
Wan Animate 2
Transfer the motion, expression, and timing of a driving clip onto a reference character while preserving identity, clothing, and the reference composition.
Kia ora. Your scale-to-zero voice model is ready.
Text + voice → speech
Pocket TTS
Fast 100M-parameter speech synthesis with 26 voices across six languages and optional authorized voice cloning from a sample.
Hi, let's think through the best things in the world.
Text + reference → expressive speech
app.nz TTS
Fish Speech synthesis with stored or inline reference voices, reusable model cache, and serverless-to-warm-pod promotion.
Text → video + native audio
MiniMax H3
MiniMax H3 text-to-video with structured audiovisual prompts, GPU AV1 WebM delivery, R2/HF weight cache, and optional EasyCache operator sweeps in the open-source adapter.
Image + prompt → video
MiniMax H3 I2V
Animate a supplied first frame while preserving identity and layout. Uses the same minimax-h3 template with first_frame conditioning and native stereo audio.
Portrait + uploaded audio → talking video
MiniMax H3 Talk
Drive mouth timing and delivery from an uploaded speech clip via H3 Ref2VA. Upload a face (first_frame) plus 2–15s of talking audio; duration snaps to the clip.
Ready to ship your own?
Pick a template in Cog Studio, or register any compliant image once and reuse it everywhere.