Replicate Select
Kuratierte Replicate-Modelle als ComfyUI-Nodes (Bild, Video, Upscale, TTS, Lipsync) + Custom-Node für beliebige Modell-IDs.
Nodes (22)
Seedance 2.0 Fast — ByteDance's video model, audio included, bill included
Seedream 4.5 in ComfyUI — the ByteDance flagship you can't download
Paying to run Real-ESRGAN is usually silly — unless you've got no GPU
SUPIR without the 16GB VRAM bill — but the faces still get rewritten
Replicate Custom — the escape hatch for the other 400,000 models
ElevenLabs v3 — the TTS everyone else is measured against, in your graph
Nano Banana Pro in ComfyUI — the Gemini image model, filter included
Veo 3.1 — the closed video bar, and the native-audio thing nothing open matches
HeyGen Lipsync Precision — swap a voice in existing footage, cleanly
Inworld TTS 2 — write '[whisper excitedly]' and it just obeys
Kling v3 video in your graph — including the multi-shot JSON mode
LTX-2 Distilled — the speed king of open video, rented for no GPU
MiniMax Speech 2.8 HD — the TTS with an emotion dial and pause markers
Pruna Video Avatar — a photo, a script, and a talking head walks out
Z-Image Turbo — the open 6B model the community runs, rented per call
Z-Image Turbo I2I — img2img with LoRA support, all in the cloud
Qwen Image Edit Plus — instruction-based image editing, cloud-style
Sync Lipsync 2 Pro — video-to-video lipsync with an actual mismatch mode
Prompt (Replicate Select) — the world's most honest node
VEED Fabric 1.0 — the three-input talking head
Turn a still image into a Wan 2.2 video without a GPU or a 27GB download
Type a prompt, get a Wan 2.2 video — no weights, no VRAM, just a token
ComfyUI-Replicate-Select
Kuratierte Replicate-Modelle als ComfyUI-Nodes — schlank statt Katalog-Flut.
Architektur nach dem Vorbild von ComfyUI-Runware:
nodes.json (Katalog) + Generator + ReplicateCustom als Escape-Hatch.
Modelle (20)
| Kategorie | Nodes | |---|---| | image | Seedream 4.5 · Nano Banana Pro · Qwen Image Edit Plus · Z-Image Turbo · Z-Image Turbo I2I | | video | Kling v3 Video · Veo 3.1 · Seedance 2.0 Fast · LTX-2 Distilled · Wan 2.2 T2V Fast · Wan 2.2 I2V Fast | | upscale | SUPIR · Real-ESRGAN | | tts | MiniMax Speech 2.8 HD · Inworld TTS 2 · ElevenLabs v3 | | lipsync | VEED Fabric 1.0 · Pruna Video Avatar · Sync Lipsync 2 Pro · HeyGen Lipsync Precision |
Dazu: Replicate Custom — beliebige Modell-ID + JSON-Input.
Auth
REPLICATE_API_TOKEN als Umgebungsvariable (auf Modal: Secret mcp-keys).
Kein Key im Node, keine Key-Dateien.
Outputs
- image → IMAGE-Tensor (direkt weiterverdrahtbar)
- audio → AUDIO (ComfyUI-nativ, z.B. → SaveAudio oder → Lipsync-Nodes via Upload)
- video → MP4 im Output-Ordner, Node gibt den Pfad als STRING zurück
Katalog aktualisieren / Modelle hinzufügen
# Modell in scripts/build_catalog.py → META eintragen, dann:
python scripts/build_catalog.py
Zieht die Input-Schemas von den öffentlichen Replicate-Schema-Seiten — kein Token nötig.
Hinweise
- Bild-Inputs werden als PNG hochgeladen (klein → data-URI, groß → Replicate Files-API)
- Audio-Inputs (Lipsync) akzeptieren ComfyUI-AUDIO und werden als WAV hochgeladen
- Video-Inputs erwarten eine URL (STRING) — Replicate braucht abrufbare Dateien