NOMAD Flux 2 Klein 9B
Maximum fidelity and photographic detail. Ideal for background swaps, photo restoration, and pristine native-resolution finishes.
Native resolution · Maximum detail
Engines & controlsYour vision.
Image synthesis & editing
Intelligent assistance & prompts
Reference image understanding
Your vision.
Your choice of engine.
Create, edit and refine with a growing ecosystem of image models. Choose the engine for the task, then shape the result in a familiar workspace.
Generation
Create images from text, edit targeted areas, and refine finishes with full control on your hardware.
Text
Expand brief ideas into complete scenes, direct compositions, and structure parameters with zero cloud.
Vision
Read input photos, analyze artistic details, and extract palettes or styles to guide new creations.
What runs
The engines behind the flows.
Each one has a clear purpose and unlocks a different creative possibility.
Generation & editing
Flux 2 Klein 4B Turbo
Fast and ultra-lightweight. Perfect for testing ideas on the fly and sketching in seconds without stressing your GPU.
Fast & lightweight · Sketches in seconds
Qwen Image 2.1
Generates from text, refines portraits, and edits specific parts of the image with ease.
Text-to-Image · Retouch · Edits
Boogu Image Edit Turbo
Apply changes and edits in seconds. Direct, zero-delay image modifications for agile workflows.
4 steps · Instant response
Z-Image Turbo · Anima · HiDream
Explore diverse visual aesthetics, from anime and illustration to conceptual looks with strong personality.
Illustration · Anime · Visual signature
Multi Composer
Combines 2 to 4 images into a single cohesive scene using spatial color zoning before harmonious generation.
2 to 4 images · Color zoning
Selection & matting
Meta SAM3
Describe what object to isolate and get an exact, automatic cutout mask powered by text.
Isolate objects · Exact mask
BiRefNet
Surgical background removal that preserves fine hair strands, transparency, and delicate edges.
Hair edges · Transparent cutout
RMBG-2.0
Fast, consistent background isolation for product photography, e-commerce, fashion, and everyday assets.
Clean silhouettes · Catalog & photo
Style & resolution
ByteDance USO
Transform scenes and environments while keeping your subject's exact facial identity intact.
Locked identity · New environment
RF Inversion
Transfer color palettes, lighting, and mood from any reference image directly to your artwork.
Match lighting & atmosphere
Krea 2 Turbo
Instantly transfer the vibe, texture, and aesthetic of an inspiration image to your render.
Visual vibe · 1 inspiration image
SeedVR2 · RealESRGAN
Upscale resolution with crisp fidelity, recovering real textures without hallucinations.
Crisp details · Zero artifacts
Portrait Relight
Reposition key and rim lighting in 3D across any portrait or render while preserving subject identity.
Studio light · Reposition light
Pixel Matcher
Local CPU algorithm (KMeans) converting renders into authentic retro pixel art with pixel-perfect precision.
Pixel-perfect · 100% local CPU
Local infrastructure
ComfyUI
The execution engine running locally on your hardware. 100% private, zero queues, and unlimited offline renders.
100% private · Zero render costs
LM Studio
Local intelligent assistant that refines prompts, describes scenes, and recommends styles without cloud queries.
Prompt assistant · Local vision
CivitAI Integration
Paste CivitAI generation blocks. Automatically maps base models, LoRAs, and sampling settings for local execution.
CivitAI to local · Zero setup
Quality & speed modes
💡 Nomad Explains Simply
How does Nomad tune precision for your GPU?
Zero technical jargon! Choose your priority for generating images.
⚡ Fast
Quickly experiment with ideas without stressing your GPU.
w4a8
·
int4
⚖️ Balanced
Near-original visual fidelity rendered in just seconds.
int8
·
fp8
💎 Max
For the final render where every pore, light reflection, and texture matters.
bf16
If a workflow declares no quantisation block, Nomad simply runs its default checkpoint. Missing tiers are skipped.
Language models
One more model.
One more model.
This one writes, it doesn't draw.
Several Create flows produce nothing until a language model has written something first. Assist expands your one line into a full scene. Scenes writes a series that hangs together. Layout emits the JSON that places every element on the canvas. That model runs in LM Studio, on your machine, beside everything else — and which one you pick changes how fast the app feels and how different two runs of the same idea come out.
Qwythos 9B Q4
The recommended default for Assist expansions, rich scene direction, and creative dialogue.
5.2 GB disk · 8 GB VRAM
Ornith 1.5 9B Q4
Full 9B creative versatility on the smallest VRAM footprint, leaving maximum room for image models.
5.4 GB disk · 8 GB VRAM
Gemma 4 12B QAT
Engineered for strict JSON coordinates, storyboard panel grids, and precise visual layouts.
6.5 GB disk · 12 GB VRAM
Ministral 3 14B Reasoning
Deliberates step by step before outputting. Ideal for complex character logic and strict world rules.
7.7 GB disk · 12 GB VRAM
Qwen 3.8 27B Ridge 3.7bpw
High-density prose and subtle atmospheric nuance, tuned for workstations with 16 GB+ VRAM.
11.7 GB disk · 16 GB VRAM
Qwen 3.8 27B Full Q4_K_M
Full uncompressed precision with maximum vocabulary depth for dedicated 24 GB workstations.
15.7 GB disk · 24 GB VRAM
View full benchmark & memory table
| Model | On disk | Comfortable on | What it's good at | Speed | Thinking |
|---|---|---|---|---|---|
qwythos-9b Q4 |
5.2 GB | 8 GB | Everything. The one we'd default to | Fast | Can be switched off |
ornith-1.5-9b Q4 |
5.4 GB | 8 GB | The same, on the smallest memory footprint we measured | Fast | Can be switched off |
gemma-4-12b QAT |
6.5 GB | 12 GB | Strict JSON and templates. Repeats itself more between runs | Fast | Can be switched off |
ministral-3-14b reasoning |
7.7 GB | 12 GB | Only if you specifically want a model that thinks first | Slow on batches | Always on |
qwen3.8-27b 3.7bpw |
11.7 GB | 16 GB | Nothing the 9B doesn't already do faster | Slow | Can be switched off |
qwen3.8-27b Q4_K_M |
15.7 GB | 24 GB | Won't fit beside your desktop on a 16 GB card | Spills to CPU | — |
Measured September 2026 on a 16 GB card with a browser and an editor open, across the four kinds of request the flows actually make. Speed is relative on purpose: your card, and what else you have running, move the absolute numbers more than the model does.
In short:
Two things nobody tells you first time
Bigger is not better here
Don't grab the largest model. A 27B takes four times longer and yields less variety than a 9B. For prompt expansion and scenes, a 9B is faster and more creative.
"Reasoning" in the name is a warning label
Reasoning models think before answering and multiply generation time by five without improving output quality in these flows.
Vision models
And a third model.
And a third model.
This one only looks.
Drop an image into Nomad Studio and something has to read it. Describe Image turns a picture into a prompt. Bulk Vision writes one prompt per image across a whole folder. Character and style references, and the Layout scene builder, all start by looking at what you gave them. That job goes to a vision model — a separate setting from the one that writes, and the one people forget to choose.
Gemma 4 12B QAT
The recommended default for vision. Excels at describing complex scenes at the exact length requested.
6.7 GB disk · 12 GB VRAM
Ministral 3 14B Reasoning
Delivers the deepest and most nuanced visual reads in our benchmarks, with zero lag on modern GPUs.
8.5 GB disk · 12 GB VRAM
Qwythos 9B Q4
Already your creative writing engine. Point vision here too for rock-solid reads with zero extra downloads.
6.1 GB disk · 8 GB VRAM
Ornith 1.5 9B Q4
Matches the 12B in read accuracy on the smallest memory footprint. Leaves maximum headroom for image renders.
6.2 GB disk · 8 GB VRAM
Qwen 3.6 27B MTP · Q3
Massive linguistic depth for batch analysis and complex visual descriptions on 16 GB+ cards.
11.7 GB disk · 16 GB VRAM
Qwen 3.5 4B MTP
Featherweight option for rapid preliminary tagging on entry-level GPUs when memory is severely tight.
3.1 GB disk · 6 GB VRAM
View full vision benchmark & memory table
| Model | On disk | Comfortable on | What it's good at | Hard image | Speed |
|---|---|---|---|---|---|
gemma-4-12b QAT |
6.7 GB | 12 GB | The one we'd default to. Best at writing the length you asked for | Reliable | Fast |
ministral-3-14b reasoning |
8.5 GB | 12 GB | The most detailed reads of the whole table, and no slower for it | Reliable | Fast |
qwythos-9b Q4 |
6.1 GB | 8 GB | Already your writing model. One model, both jobs | Reliable | Fast |
ornith-1.5-9b Q4 |
6.2 GB | 8 GB | The same, on the smallest footprint that still never missed | Reliable | Fast |
qwen3.6-27b MTP · Q3 |
11.7 GB | 16 GB | Nothing the 12B doesn't do in half the time | Reliable | Slow |
gemma-4-e4b Q4 |
5.9 GB | 8 GB | Describes an image well, then goes blank when asked to copy a style | Fails | Fast |
qwen3.5-4b MTP |
3.1 GB | 6 GB | Works, until it doesn't, with nothing on screen to tell you which | Coin flip | Very fast |
Measured September 2026 through the app itself, not against the model directly. Three references of rising difficulty — a face filling the frame, a wide rooftop scene with a small figure, a cluttered interior — across both jobs the app asks of a vision model: describe this, and describe this in the style of that. "Hard image" is the wide scene with a style to copy, which is where the differences show. Reliable means three good reads out of three. "Fails" means three out of three came back as nine words of style and no description at all. The coin flip is exactly that: four usable reads in eight attempts.
How it understands a workflow it has never seen
Bring your workflow.
Bring your workflow.
Make it feel familiar.
Import a ComfyUI workflow and expose supported controls in the studio. Available controls depend on its nodes; its models and dependencies must be installed in your environment.
Finds the parts
Positive and negative prompt nodes, samplers, seed and noise inputs, latent nodes, guiders, checkpoint and UNET loaders, CLIP last-layer nodes, quality presets, image loaders and LoRA loaders.
Rebuilds subgraphs
ComfyUI's API export flattens subgraphs. The studio reconstructs their boundaries from the parentId:childId convention — and when exactly one external input of the same type can replace a subgraph's output, it marks that subgraph as an optional on/off toggle with the rewiring precomputed.
Chains LoRAs
LoRAs are injected into an imported workflow through a sentinel key, normalised and extracted back out again — with a catalogue that pulls identifiers, details and previews from CivitAI.
Verifies before running
Model files are checked on disk across every candidate folder before a run starts, and missing files can be downloaded from the import screen.
Detects upscale style
An imported upscale workflow is inspected for ImageScaleBy versus ImageScale, and the right parameter is exposed automatically.
Benchmarks vision
A built-in batch benchmark compares quality and speed of "describe image" across vision models, with progress polling and a downloadable report.
Bring your own workflow.
Nothing here is a closed catalogue. Import a checkpoint or a workflow and it becomes a first-class flow with the same controls as everything else.