← Back to AI Hub

🎬 Using Wyltek Studio

A local-first creative suite for image, audio, and video generation — everything running on your own GPU, no API keys for the core workflows. This guide covers install, configuration, and the small number of concepts that separate "it's generating garbage" from "that looks like a finished asset."

1. What Wyltek Studio is (and isn't)

Wyltek Studio is a local generative-AI creative suite built around ComfyUI and a small set of purpose-built studio pages. The single Python server coordinates:

• Image generation — 15+ models (SDXL, Flux dev/schnell, SD 1.5, and full-res photoreal/artistic checkpoints) with LoRA support and compare mode
• Prompt enhancement — local Ollama models transform crude prompts into detailed SDXL-ready input
• Studio tools — frame extraction, background removal, SAM click-to-segment, TTS, voice cloning, music generation, and video via AnimateDiff
• Projects — shared timeline/asset system so you can assemble generations across tools
• Auto-scoring — CLIP aesthetic scores on every image so you can sort by quality

It isn't a cloud SaaS, and it isn't a thin wrapper around DALL-E or Midjourney. Every generative step runs on your hardware. The trade-off is you need a capable GPU (8 GB+ VRAM minimum for SDXL) and a few gigabytes of disk for checkpoints.

2. Quick-start install

git clone https://github.com/toastmanAu/wyltek-studio ~/open-palette
cd ~/open-palette
./install.sh

The install script walks you through:

1. Python venv + dependencies
2. ComfyUI install (if not already present)
3. Ollama detection + default model pull (optional — only if you want the Prompt Optimizer)
4. Baseline model catalog — SDXL base, Flux Schnell, LoRAs, VAEs

Then launch:

cd ~/open-palette
./run.sh    # or: python server.py

Studio opens at http://localhost:7860. Config lives at config.yaml in the project root.

3. Feature map

4. Image generation — choosing a model

Wyltek Studio's model catalog (Settings → Models) lists what's available and what's installed. Pick based on intent:

Intent	Model recommendation	Why
Photoreal portraits / architecture	RealVisXL V5.0	Heavily finetuned photo SDXL. Excellent skin tones, interiors, landscapes.
Versatile "all-rounder"	juggernautXL v9	Photoreal-leaning but comfortable with illustration. Good with LoRAs.
Illustration / fantasy	DreamShaper XL Turbo	Only needs 6–10 steps (Turbo distilled). Great for concept art iteration.
Neutral "canvas" for LoRA testing	SDXL base 1.0 (Q4)	Doesn't fight LoRAs the way heavily-finetuned bases do. Best for evaluating new LoRAs.
State-of-the-art quality	Flux dev or FLUX.2-klein	Best prompt following and coherence. Slow + heavy (16 GB VRAM class).
Fast iteration	Flux Schnell or SDXL + Lightning LoRA	4-step distilled. Draft-quality previews in seconds.

5. Prompt craft & the Prompt Optimizer

The single biggest lever on image quality isn't the model — it's the prompt. "A cat" gives SDXL nothing; "A detailed, whimsical cat character, vibrant colors, professional illustration, trending on ArtStation" gives it a direction. The Prompt Optimizer (OP my prompt) does that transformation via a local Ollama LLM.

The dropdown next to the OP button lets you swap Ollama models per call. Your selection persists in localStorage across sessions; the server default (configured in Settings → Prompt Optimizer) is the fallback.

6. LoRAs — styling your output

Style LoRAs transform the aesthetic of your output — voxel, pixel art, watercolour, crayon, etc. Wyltek Studio ships with six style LoRAs preconfigured. Three things matter when using them:

1. Pick one that matches your intent — voxel for game-ready 3D, crayon for hand-drawn illustration, pixel-art for retro-game sprites, and so on.
2. Trigger words — each LoRA has activator tokens. "voxel art" for voxel-xl, "crayon drawing" for crayon-style-xl. Without them, the LoRA barely fires. Wyltek Studio auto-injects these for you when your prompt is bare.
3. Strength — model strength ~0.8–1.2 and clip strength ~0.5–0.9 are the usable range. Below 0.7 you get a faint tint; above 1.2 images burn.

7. Projects — assembling outputs into timelines

Every image, audio clip, and video frame you generate lands in Wyltek Studio's Projects system. Each project has a shared assets folder and a timeline where you arrange generations in sequence for export. The flow:

1. Create or pick a project from the Projects page.
2. Generate content in any studio tool — outputs auto-save into the project.
3. Drag assets onto the timeline, adjust durations, compose audio / video / image layers.
4. Render the composed timeline as a video file.

This is what separates Wyltek Studio from a loose collection of AI tools — you can iterate on a music track, generate matching frames, run TTS for narration, and assemble them all in one place without leaving the app.

8. Claude Code integration

Wyltek Studio ships several skills for Claude Code that treat the studio as a callable service. Relevant ones:

• /cut-subject — background-removal skill that wraps Wyltek Studio's rembg + SAM endpoints. Run /cut-subject ~/photos/x.jpg and Claude picks the right model, calls Wyltek Studio, and returns the transparent PNG.
• /ingest-doc — extracts text / frames / audio from any document and pushes it through Wyltek Studio's processing pipeline.
• /quote-meme — quote-card meme generator using Wyltek Studio's image tools.

See the Skills guide for the full list.