What Vision Studio does, grouped the way you work: generate, edit, animate, organize, source models, run graphs, tune the engine, and choose where each job runs. Each entry names the files in the public repository that implement it - open any SRC chip to read the code behind the claim.
Image and video models that run on your own hardware, and the controls that shape a render: sampler, steps, guidance, seed, framing, LoRAs, ControlNet and reference images.
FLUX.1 [dev] and [schnell], Stable Diffusion 3.5 Large and Medium, Stable Diffusion XL and Stable Diffusion 1.5, each run through its Hugging Face Diffusers pipeline.
LTX Video and AnimateDiff generate a clip from a prompt; Stable Video Diffusion animates a still image you supply. Clips are written as MP4.
Prompt and negative prompt, eight samplers (Euler, Euler a, DPM++ 2M, DPM++ 2M Karras, DPM++ SDE, DPM++ SDE Karras, DDIM, UniPC), 1 to 50 steps, CFG from 1 to 20 and a fixed or random seed.
Stack any number of installed LoRAs, each with its own weight from 0 to 2. They apply to image generation and to AnimateDiff and LTX video, and are removed again after every job.
Guide composition with edge, depth, pose, scribble and normal-map layers placed on the canvas.
Two or more reference layers condition a render through IP-Adapter: region-by-region with masks on SD 1.5 and SDXL, whole-image on FLUX.1 [dev]. A single reference layer runs as image-to-image.
Queue a list of prompts at once: each prompt becomes its own job, four submitted at a time, and one click turns a prompt into four styled variations.
Studio runs show the image forming, decoded from the latents every half second by Tiny-VAE decoders that ship inside the app. Every job reports its progress over a local WebSocket.
Model-backed tools that change the pixels, run by the local engine on the image you are working on, plus the canvas layers that steer the next render.
Background removal (U2-Net), AI upscale (Real-ESRGAN, general or anime, 2x or 4x), face restoration (GFPGAN v1.4), style transfer, generative fill, object removal, AI expand and background replace.
Paint a mask to regenerate part of an image, or extend the canvas outward, with your selected checkpoint. FLUX inpainting switches to FLUX.1 Fill.
Inpaint masks, ControlNet guides and reference images live as layers on the canvas, and the next render reads them.
Geometric edits are applied by the engine and saved as a new PNG in its own output folder; the original file is left as it was.
Plan a sequence as scenes, then cut generated clips together and export a finished MP4.
Lay a project out as scenes and step through them in playback. Onion skin ghosts the neighbouring scenes over the active one, with adjustable count, opacity and direction (previous, next or both).
Arrange clips on video and audio tracks, split, duplicate and set a play range, then export an MP4 with cut, fade or dissolve transitions and your audio encoded as AAC.
Eight named starting points (YouTube Thumbnail, TikTok/Reels, Instagram Post, Stories, Cinematic Wide, Product Showcase, AI Portrait and 4K Wallpaper) that set the model, prompt scaffold, steps and CFG for a new project.
Everything you generate lands in a local library you can search, branch from and export.
Search by name, prompt or model, filter images, video and audio, star favourites, import files, and export one asset or many to any folder.
Every render becomes a node. Fork one to branch from its prompt, model, sampler, steps and seed, or re-roll it with a fresh seed; each child records a settings diff against its parent, and the tree survives restarts.
Analyze tags each asset from the words in its prompt (style, subject, colour and mood) with no model and no upload. The negative prompt is ignored.
Find, vet and install models without leaving the app, with the risky formats locked behind an explicit consent step.
Search both hubs from inside the app and install straight into your local library. Each result shows its tier and licence; CivitAI adult results stay hidden unless you opt in.
Pickle-format weights and models that ask for trust_remote_code are denied by default. Nothing downloads until you consent for that model, remote code is never loaded, and pickle files are converted with PyTorch's weights-only loader before use.
Each model is sized against your GPU and system memory and labelled as fitting, fitting with CPU offload, over budget, or CPU-only.
On first launch you can install a curated set of 33 models (about 137 GB) in one step, or skip it and add models one at a time. FLUX.1 [dev] and its companions are licensed for non-commercial use and are left out of the set.
Downloads run two at a time with live status. Direct links are accepted only over HTTPS from CivitAI and GitHub and are checked against a SHA-256 hash.
Bring ComfyUI graphs in, or let a ComfyUI server you already run do the rendering.
Import a ComfyUI API-format graph and run it on Vision Studio's own engine. A graph with a single KSampler becomes a normal generation carrying its prompt, negative prompt, model, steps, CFG, seed and LoRAs.
If a ComfyUI server answers at 127.0.0.1:8188 when Vision Studio starts, or at the address in the COMFYUI_URL environment variable, plain image jobs and all video jobs are sent to it instead of the built-in engine.
A local Python service does the generating on your hardware, and a Performance panel decides how hard it pushes.
A bundled FastAPI service runs PyTorch 2.5.1 and Hugging Face Diffusers on 127.0.0.1:8000, reachable only from your own machine, and its generation API requires a token the app creates at launch. It uses the first NVIDIA GPU through CUDA, Apple Silicon through Metal (MPS), or the CPU when there is no GPU.
Auto, On or Off for SDPA attention, channels-last memory, torch.compile, quantization, attention slicing and TensorRT, behind one master switch. Auto decides per model and GPU, and the panel shows what was applied, skipped or fell back.
A model that only fits with offload loads with CPU offload and VAE tiling, and an out-of-memory error steps down a fallback ladder instead of failing outright. Precision follows the hardware: bf16 where the GPU supports it, otherwise fp16 (fp32 for FLUX and SD 3.5, which are not run in fp16), and fp32 on the CPU.
Local by default. A cloud provider only when you add your own key and choose it, one job type at a time.
Prompt tools, still images and video all start on the Local route. The interface is only allowed to connect to the engine on your own machine, and the code contains no analytics, telemetry or crash reporting.
Add an OpenRouter or Hugging Face key to route prompt tools and still images to a hosted model, or video to Hugging Face. Keys are encrypted with your operating system's secure storage, and a job too large for your GPU asks before it switches route.
Prompt text and generation settings, never your images: a request that carries a mask, reference image, ControlNet guide or source image is refused on a cloud route. Cloud prompt tools can also include short excerpts of your earlier prompts as context, controlled by the AI Director setting.
With every route on Local, the app's code goes online for model searches and downloads you start and for the update check, which asks updates.vision-studio-x.com for a newer release 15 seconds after launch and every four hours. A saved OpenRouter key also lets Settings fetch that account's model list.
On the Local route, enhancement adds composition and style phrases in five modes (clarify, cinematic, concise, expand, variations) and negative-prompt suggestions come from keyword rules; no language model runs. Enhancement written by a language model needs an OpenRouter or Hugging Face key.
Eight panels, reachable from the keyboard, with desktop notifications when work finishes.
Generate (with Quick, Batch and Studio modes), Canvas, Story, Workflows, Assets, Collections, Foundry and Settings.
A skip link, arrow-key navigation across the main tabs, and keyboard-resizable panels that report their size to screen readers. An axe-core WCAG 2 A/AA check runs against the Generate and Settings panels.
A desktop notification when a generation finishes or fails, and for model downloads, each switchable in Settings.