NOMAD 
8 GB VRAM
For step-by-step tasks and quick testing. Rely exclusively on compact models (4B to 9B in Q4) to keep memory headroom clear.
4B – 9B Q4 models · Single model loaded
ExtendMake the studio
Make the studio
your own.
Bring in the workflows you love, add useful tools and connect an assistant when you need one. Extend Nomad Studio around the way you create.
Your workflow.
A familiar workspace.
Already have a ComfyUI workflow? Bring it into Nomad Studio and use its supported controls in a familiar panel.
Available controls depend on the workflow. Its models and required components must be installed in your ComfyUI environment.
Import details
Bring your own
Import any workflow. It becomes a real flow.
Drop in a ComfyUI graph and the analyser reads it — prompt nodes, samplers, seed and noise inputs, latents, guiders, checkpoint and UNET loaders, CLIP last-layer nodes, image loaders, LoRA loaders. You get a parameter panel, not a text box full of JSON.
- Verified before it runs. Model files are checked on disk across every candidate folder, and missing ones can be downloaded from the import screen.
- Annotated. Record downloads, LoRAs, optional subgraphs, requirements and persisted values for basic and advanced fields.
- Inspected without importing. The workflow editor analyses a raw JSON file and shows its parameter options before you commit.
- Branded. Upload a custom logo per imported workflow so your own graphs look like they belong.
Subgraph reconstruction
ComfyUI's API export flattens subgraphs. Nomad Studio infers their boundaries back from the "parentId:childId" ID convention. When exactly one external input of the same type can stand in for a subgraph's output, that subgraph becomes an automatic on/off toggle — an "optional part" — with the rewiring precomputed.
LoRA chains
LoRAs are inserted into an imported graph through a sentinel key (<unet_node_id>.__insert_lora), then normalised and extracted again from the UI's LoRA tree — with a catalogue that pulls identifiers, details and previews from CivitAI.
Hardware & VRAM Guide
GPU recommendations.
GPU recommendations.
Pick the right model size.
If a workflow runs in your ComfyUI, it runs in Nomad Studio: the studio uses the same GPU and the same models. These are rough guides by card memory.
16 GB VRAM
The golden standard for the full studio. Run your image generator and 9B assistant models side by side with zero lag.
Generator + 9B assistant
24 GB+ VRAM
For intensive production: huge ComfyUI graphs, dense 14B to 27B models, multi-LoRA stacks, and ultra-high resolutions.
Workstation · 14B – 27B models · Heavy batches
The rule: if it works in ComfyUI, it works in Nomad Studio. Smaller or quantised models leave more room to run an assistant alongside the generator.
Small additions.
New possibilities.
Turn an image into pixel art, remove a coloured background or combine images. Add-ons bring more ways to work into your studio.
Pixel Matcher
Turn images into pixel art.
Chroma Key
Remove a background by colour.
Canvas Composer
Combine and position images on a canvas.
Installation & development details
Add-ons
Drop a folder in. Restart. It is live.
The add-on system is modelled on Blender's: each one lives in its own folder with its own Flask blueprint, service layer and flow definition. Install, remove or update without touching a single line of core code.
- Declare and register. An add-on exposes ADDON_INFO and a register(app) hook; anything under addons/ is discovered at startup.
- Add flow steps. An optional flow.json is merged into the core catalogue, so an add-on's feature appears as a normal flow in the normal UI.
- Any technology. Pure Python, an external local API, a new model, or an entirely new UI flow — there is no capability restriction.
- Documented and open. Anyone can build one, share the folder, and have it work on any installation.
Shipping in the box
PX
Pixel Matcher
Any image to pixel art via KMeans colour quantisation. Auto-detects grid size, reduces the palette, scales up to 8×. No GPU, no ComfyUI.
CK
Chroma Key
Remove a solid or vignetted background by colour, with despeckle, feathering and edge decontamination.
DF
Edge Defringe
Strips residual background-colour bleed from the semi-transparent edge of a cutout without altering the alpha shape.
CC
Canvas Composer
Composites a movable, scalable overlay onto a background image on a real canvas.
All four are 100% local NumPy/SciPy — they run with ComfyUI switched off.
Say what you imagine.
Then make it yours.
Connect Claude Desktop, ChatGPT, Antigravity or Bionic, describe your idea and ask for changes. Nomad Studio runs the image work on your local setup.
Image processing runs locally. The connected assistant may use a cloud service, depending on your setup.
“Make me a photo of a ceramic artist at her wheel, in a workshop with the sun coming in.”
“Now make the bowl deep blue and glossy. Leave everything else alone.”
“Pull back a bit, I want to see more of the workshop.”
Three turns of one real conversation. The third one never named a file — “pull back” was enough, because the assistant is still holding the picture it just made for you.
What you can ask for
The assistant fills these in from your own words — “wide, best quality, with Monaco” is enough. You can also set them yourself.
For example
- “Make it wide, best quality, with Monaco” → the assistant sends aspect_ratio wide, speed max, model monaco
- “Same picture, but with no people in it” → it reuses the seed of the last image and adds negative people
- a lighthouse at dusk --ar 16:9 --q max --no people → you can also type the shortcuts yourself, at the end of the prompt
| Parameter | Shortcut | Values | If left out |
|---|---|---|---|
| prompt | — | Description of the image, ideally in English. Required. | — |
| aspect_ratio | --ar 16:9 | square · landscape · portrait · wide · tall | Reuses your last choice in the conversation, or asks you |
| speed | --q max · --draft | fast · balanced · max | Reuses your last choice in the conversation, or asks you |
| model | --v monaco · --anime | Any model you have installed, by its short name or part of its name | Reuses your last choice, or asks; with only one model installed, uses it |
| quantization | --quant int8 | The variants your model offers (fp8, int8, bf16…) | Picked from speed |
| steps | --steps 20 | Steps, 1 to 150 | The model's own value |
| width / height | — | Exact pixels, both together, rounded to multiples of 64 | The size of aspect_ratio |
| negative | --no people, text | What must not appear | The model's default negative |
| seed | --seed 1234 | A number, to repeat or vary an image | Random |
| count | --r 4 | 1 to 4 images per request | 1 |
Typed at the end of the prompt. They win over everything else.
Everything else the assistant can do
| Tool | What it does | Main parameters |
|---|---|---|
| Refine | Edits one image with any Refine flow you have installed: relight, remove the background, upscale, retouch, colour looks, change the text inside the image… | action, instruction, regions (marked zones); for outpaint: position, scale, expand_percent |
| Mix | Combines several images into one: a subject onto a background, several references into one scene, the style of one image onto another | action, images (each in its slot), instruction, aspect_ratio, seed |
| Sessions | Lists and reads your sessions and saves what it makes as new steps; from then on every image lands there by itself | session_id or a new name, keep_adding |
| List models and actions | Tells the assistant what you actually have installed, so it never invents a name | — |
| Last generation | Recalls the prompt, seed, model and size of the latest image | — |
Which Refine and Mix flows appear depends on the packs you have installed.
For those who want to go deeper
Technical references are here when you need them. You do not need to read them to explore what Extend offers.
Flow building reference
The flow engine
A few step types. As many flows as you install.
A flow is a JSON file listing steps. A new step type becomes available to every flow at once, so the studio grows with the packs you add without the interface fragmenting.
| Step type | What it does |
|---|---|
| workflow_picker | Choose among your active imported workflows; auto-selects and skips itself when there is only one |
| prompt_input | Text entry, optionally pre-filled from an earlier step, with image drop and loop-back for iterative flows |
| param_group | Exposes seed, steps, CFG, image count, width, height, aspect ratio and denoise — the flow decides which are visible |
| json_input | Edits a structured variable template key by key, when one text prompt is not the right shape |
| llm_transform | Sends a previous step to the language model with an instruction; the result is editable before it continues |
| llm_chat | An open conversation where the model interviews you and builds the prompt incrementally |
| style_picker | The style gallery, usable as a step in any flow — not only in Create |
| character_picker | Chooses a saved character and passes its activation keyword downstream |
| image_picker | Pick from the album, upload, or drag in — for compose, upscale and image-to-image |
| vision_describe | Runs a vision model over a reference image and pre-fills a linked prompt step with what it saw |
| workflow_params | The advanced panel: fields mapped to ComfyUI node inputs in nodeId.inputKey dot notation |
| action | Calls an app action and shows the result — the terminal step of most flows |
| result_view | Shows the output with save, upscale, send-to-refine, regenerate, re-enhance and story continuation actions |
Assistant connection details
Under the hood
What that connection actually gives you.
A built-in MCP server exposes the studio over stdio to any MCP client — Claude Desktop, ChatGPT, Antigravity and Bionic all connect today. It is not a stripped-down side channel: the assistant can create, refine and mix images and work with your sessions, using the same models, quantisation options and speed presets as the app.
Every Refine and Mix flow you install is reachable from a conversation — background removal, outpainting, relighting, object removal, colour grading, portrait retouching and more. The catalogue is read from your installed flows at the moment you ask, so a pack you install this afternoon is available in the chat immediately. are reachable from a conversation today — background removal, outpainting, relighting, object removal, watermark removal, colour grading, portrait retouching and the whole preset library. That catalogue is read from your own installed flows at the moment you ask, so a pack you install this afternoon is available in the chat immediately.
- Its own identity. Every agent registers as a viewer with a name, slug and ID, so its generations are attributed and get their own album.
- Remembers your preferences. Aspect ratio and speed carry across a conversation instead of being asked again each time.
- Asks rather than guesses. When a preference genuinely is not known, the tool returns an explicit instruction for the model to ask you — instead of silently picking values you did not choose.
- Reads from disk. Because the agent runs on the same machine, it loads the finished image directly and returns an inline thumbnail plus a direct URL, skipping the session cookie system entirely.
- Keeps hold of the picture. Edits chain without naming a file — say “now remove the background” and it works on what it just made for you.
- Negotiates capabilities. Behaviour adapts to what the connected MCP client actually supports.
- Repeatable results. Every answer includes the seed: the same prompt, seed, model and size give the same image.
- Lands in your sessions. Choose a session once and every image the assistant makes or edits is added as a step you can open in the app.
Profiles & privacy options
Shared machines
Profiles, PINs and a mode that leaves nothing behind.
Viewers
Pick a profile on launch, optionally behind a four-digit PIN. Histories, depots and sessions stay separate per person.
Vanish mode
A private profile whose generations pass through temporary memory and disappear — for test runs you do not want in your history or on your disk.
Agent identity
A separate identity system for MCP agents that lives alongside human profiles without mixing into them.