Pinokio

Spectrum acceleration is now available in H3 Studio!Enable it under Settings → Video engine → Spectrum . It’s optional, experimental, and requires no additional LoRA. In our matched 608p test, total render time dropped from 26:15 to 16:06 — 38.65% less time . Recent 1088p clips rendered in around 35 minutes . Compared with my previously observed roughly one-hour renders, that suggests about 42% less time , although this 1088p comparison is approximate. My system: RTX 5060 Ti 16 GB, Ryzen 5 5600X, 32 GB RAM, Windows. Results vary by workload. Spectrum can affect motion and speech, so test a short clip first. H3 Studio now offers three video engines: - MiniMax H3 — the standard engine. - H3 Singularity — an alternative H3 checkpoint. - H3 Hybrid — an experimental H3 checkpoint. Switch between them under Settings → Video engine . Optional engines require their model files to be installed. The Spectrum benchmark above was measured with MiniMax H3 Ref2VA INT8. It does not establish the same speedup for Singularity or Hybrid.

Captionz

@supersoniquestudio6h ago

1-click Pinokio launcher for Captionz — batch image captioning with local Ollama vision models. What it does Installs and launches Captionz in one click: Install — clones mikecastrodemaria/Captionz into app/, creates a venv and installs the requirements (Pillow, NiceGUI, Gradio). No PyTorch: the models run in Ollama, or in a bundled llama.cpp server (see below). Start (web UI) — runs the NiceGUI interface (python app.py --ui web) and opens it. Start (Gradio UI) — runs the Gradio interface (python gradio_app.py), the same one used on Hugging Face Spaces, here with the Ollama backend. Update — resets the launcher and the app to their origin/main, refreshes deps. Reset — removes app/ (and its venv) to reinstall from scratch. Features (Captionz) Vision-model detection · composed prompts (caption type × length × 20 checkable options × character name) or custom prompt with a live preview · sources: local path, upload, clipboard paste · image preview · editable captions · skip / overwrite / append policy · prefix / suffix (trigger word) · live status (model loading, timer, percent, ETA) · max-tokens cap and thinking switch · dark mode · CLI and benchmark script. See the app repo for full docs. Requirements Pinokio installed. Either Ollama running on the machine (or reachable on the network: change the URL in the UI) with at least one vision model, for example ollama pull qwen3-vl:8b, or no Ollama at all: pick the backend "llama.cpp (sans Ollama)" in the UI. On first captioning Captionz downloads a prebuilt llama-server for your machine (CUDA / Vulkan / CPU / Metal, 150–500 MB) and the default model Qwen2.5-VL-3B-Instruct Q4_K_M (~2.9 GB) into app/llamacpp/. Other GGUF models can be downloaded from Hugging Face or imported from a local Ollama store from the "Modèles llama.cpp…" dialog. Captions are written next to each image as image.txt. How to use Install, then Start (web UI) and click Open Web UI. Pick the vision model, add a folder path (or upload / paste images). Choose a caption type, length and options; check the final prompt preview. Click Captionner tout. Edit any caption in the right panel and save. API Captionz has no HTTP API of its own. Two programmatic entry points exist in app/: Command line (Python, inside the launcher venv) python cli.py ./dataset --recursive --type "Booru tag list" --length short python cli.py ./dataset --model qwen3-vl:8b --option 3 --option 15 --prefix "mystyle, " python cli.py --list-models python cli.py --list-options Python (import the core) import sys; sys.path.insert(0, "app") from pathlib import Path from captionz_core import Settings, Job, make_backend, caption_job s = Settings.load() s.model = "qwen3-vl:8b" s.caption_type, s.caption_length = "Training caption (paragraph)", "long" job = caption_job(Job(Path("photo.jpg")), s, make_backend(s)) print(job.status, job.caption) Gradio UI API (when "Start (Gradio UI)" is running) Gradio exposes every event as an endpoint; open <url>/?view=api for the generated docs, or from Python: from gradio_client import Client, handle_file c = Client("http://127.0.0.1:7860") print(c.view_api()) JavaScript (Gradio client) import { Client } from "@gradio/client"; const c = await Client.connect("http://127.0.0.1:7860"); console.log(await c.view_api()); Curl (Ollama directly, same request Captionz sends) curl http://localhost:11434/api/chat -d '{ "model": "qwen3-vl:8b", "stream": false, "messages": [{"role": "user", "content": "Describe this image in one short sentence.", "images": ["<base64 image>"]}] }'

Phosphene

@bizarro1d ago

Phosphene 4.19 - Last Take One night, one director, one take before the sun comes up. Every shot, every sound and the song were made in Phosphene on one Mac. - The rain on the backlot was a silent clip. Add sound wrote the rain and the thunder that hits right on the lightning, and kept every frame exactly as it was. - The stage coming alive is Keyframes: one dark frame with only the ghost light on, one fully lit frame, and Hailuo H3 made everything in between. Extend then carried it 5 more seconds. - The long push-in on Bizarro is a 5 second clip made longer twice with Extend. Each new piece carries the last frames and their sound, so it plays as one unbroken 15 second move. - The film camera rolling is Fast HD: a quick small H3 render, finished at 1280x768 by LTX's x2 Upscale & Face Fix. - The finale is Lip-sync on H3 with our own song: the song is kept exactly, only his face is made to sing it. The song itself came out of Phosphene's Music Studio and scores the whole film. Also in 4.19: - 10 and 15 second H3 shots join more smoothly (the jump in sound at the seam is mostly gone). - Fast HD: about a third quicker than High, for previews and quick drafts. - Extending a clip again and again no longer wears it down ( 48) - Characters you trained in Phosphene finally look like who you trained - thank you @tanis2000 ( 62) 4.19.1: voices trained on 4.19.0 are repaired on Update, a failed training now says why, and the Stop early button no longer covers the time left (thanks @blackest) Press Update in Pinokio, then restart. https://github.com/mrbizarro/Phosphene/releases/tag/v4.19.0

Iris-3B Studio

@thomas_eric2d ago

Iris 3B Studio is Now Available!Iris 3B Studio is Now Available! Not only is a incredible text-to-image model, but it also has upscaling capabilities and depth map generation!

SeedVR2-7B MLX

@rdx_ai3d ago

New Update with Multiple/batch import, Library tab!Updated with Multiple/batch import support, New Library tab for viewing multiple images in a session.. Fastest and most Memory Efficient Seedvr2 implementation for Apple Silicon Mac..

LightOnOCR 3

@cocktailpeanut3d ago

LightOnOCR-3: Turn PDFs and images into structured text locally (Mac, Windows, Linux)LightOnOCR 3 is now available on Pinokio. It's a sequel to the LightOnOcr 2 model, which you can try here (thanks to @theawakenone ) https://pinokio.co/apps/github-com-theawaken1-lightonocr-2-1b-pinokio This LightOnOCR 3 launcher implements the backend using huggingface transformers and lets you run 3 variations of the model: - 0.8B - 1B - 4B The launcher includes LightOn’s original web interface, available at: https://github.com/lightonai/LightOnOCR Demo Here's how it works. You just launch the app and open any PDF or image file: Features - View the source document alongside the extracted text - inspect layout boxes showing where content came from - revisit previous runs. Examples

Type:api

Platform:All

GPU:All

RadarFeed

can it create animated prompts of an uploaded image?

@arif1979 · Image to Prompt

not working at all...

@akos2 · SeedVR2-7B MLX

I think a 4-step turbo version is out? https://github.com/speridlabs/iris-3b iris-3b-turbo

@cocktailpeanut · Iris-3B Studio

Checked in to Wan2GP.

@lucassribeiro · Community

Maestro: [Runtime] FlashAttention is unavailable (No module named 'flash_attn'); continuing with SageAtt…

@ilovergb · Maestro

App: Maestro (Maestro.git) Repo: https://github.com/Blizaine/Maestro.git Generated: 2026-10-12T19:20:...

Discoveredjust now

free-video-generator

Completely free AI video generator — no GPU, no subscription, no limits. Built on Agnes AI free models.

Followed2 min

Wan2GP

Super Optimized Gradio UI for AI video creation for GPU poor machines (6GB+ VRAM). Supports Wan 2.1/2.2, Qwen, Hunyuan Video, LTX Video and Flux. https://github.com/deepbeepmeep/Wan2GP

Followed2 min

HeartMuLa Studio

A professional, Suno-like music generation studio for HeartLib. https://github.com/fspecii/HeartMuLa-Studio

Followed3 min

Maestro

An all-in-one, 100% local AI creative studio, director, and multi-track editor. Generate with MiniMax H3, LTX-2.5/2.3, Wan, Flux, Qwen, and more; turn an idea or song into a planned production; then finish it on the timeline. Requires an NVIDIA GPU (6GB+ VRAM).

Followed4 min

SongGeneration Studio

AI Song Generation with Full Style Control - Generate complete songs with lyrics, vocals, and instrumental tracks using Tencent AI Lab's SongGeneration (LeVo) model. [NVIDIA ONLY]

Apps