AI Image Generationv1.0.5

Nano ImageEdit

Nano ImageEdit is a local desktop image generator and editor that runs the Flux.1 diffusion transformer on a single NVIDIA GPU on Windows, supports text-to-image, image-to-image, and reference-image conditioning, and uses a streaming DiT path so cards with under 12 GB of VRAM can run the same 12-billion-parameter model.

Flux.1
12B-parameter diffusion transformer
<12 GB VRAM
Streaming DiT for limited-memory cards
Local-only
No subscription · no cloud generation
Single-GPU local inferenceFlux.1 — 12 B parametersNo subscription · one-time licensePrompts and outputs stay local
01Flux.1 backbone

12 billion parameters, on the user's GPU.

Nano ImageEdit runs the Flux.1 diffusion transformer (12 billion parameters) on a single local NVIDIA GPU rather than routing prompts to OpenAI, Midjourney, Adobe Firefly, or Leonardo's cloud endpoints, so prompts, references, and outputs stay on the local disk.

  • Flux.1 — 12 B-parameter diffusion transformer
  • On-device inference, no remote endpoint
  • Prompts and references never uploaded
  • Output rights retained 100% by the user
Flux.1 · open weights
02Text-to-image

Prompt to image, in one app.

Text-to-image accepts a prompt plus a small set of generation knobs — guidance, sampler steps, seed, aspect ratio — and produces an output image at a chosen size, with no subscription, no per-image charge, and no waitlist.

  • Prompt-driven generation with seed control
  • Common sampler knobs exposed (steps, guidance)
  • Aspect-ratio presets for print, social, and 3:4 portrait
  • No subscription, no per-image charge
Prompt · seed · aspect
03Image-to-image

Edit with a reference photo.

Image-to-image accepts a reference image plus a prompt and applies the prompt's edit while preserving the reference's structural layout — useful for product re-photographs, color-graded variants, and edited stock that would otherwise require Photoshop and a remask.

  • Reference-image conditioning
  • Structural layout preserved
  • Per-image strength slider
  • Useful for product re-photography and stock variants
Reference + prompt → output
04Streaming DiT

Run Flux.1 on a 12 GB GPU.

The streaming DiT path streams attention layers between disk, system RAM, and VRAM during inference, enabling cards with under 12 GB of VRAM (RTX 3060 / 4060 / 5060 class) to run a model that otherwise targets larger memory footprints.

  • Streams attention layers across the memory hierarchy
  • Runs on RTX 3060 / 4060 / 5060 class cards
  • Same Flux.1 weights as larger GPUs
  • Single in-app toggle, no manual quantisation
<12 GB VRAM · single toggle
05License protection

Product-bound license.

Each install activates against a product-bound license, which prevents redistribution of the bundled Flux.1 weights and downstream training-set leakage; license activation is the only network call the application makes during a generation session.

  • Product-bound license activation
  • Bundled weights protected from redistribution
  • License is the only network handshake
  • Inference itself is fully offline
Activation only · offline inference
06How it compares

Midjourney quality, locally, no subscription.

Midjourney, DALL-E, Adobe Firefly, Leonardo, and Krea run server-side and bill per generation or per seat. Nano ImageEdit runs Flux.1 locally on a single NVIDIA GPU with a one-time license, no per-image charge, and no requirement to send prompts to a remote endpoint.

  • Local inference on a single license
  • No per-image or per-seat charges
  • Same Flux.1 model on every supported GPU
  • No remote prompt logging
vs Midjourney · vs Firefly · vs Leonardo

Documentation

Honest scope, known limits, sources.

The marketing copy above tells you what Nano can do. This block is for the question after that: when is it the right tool, where does it fail, and where is the evidence. Updated .

Scope

Best for

  • Concept art and illustration on a 12 GB-VRAM card
  • Image-to-image edits where structure must be preserved
  • Studios that cannot ship prompts to Midjourney or DALL-E
  • Iterating offline with seed control and reproducible re-rolls
  • Replacing per-image cloud subscriptions with one-time license

Not recommended for

  • Real-time interactive painting on the same canvas
  • Apple Silicon Macs (Windows + NVIDIA only in v1.0.5)
  • Bulk e-commerce upscaling — use Nano ImageEnh Pro 3.0
  • Multi-character compositions with strict identity locks (use FacialEdit)
  • Photoreal text rendering (Flux.1 still struggles with paragraphs)
Known limitations

Boundary conditions and failure modes from internal QA. Listed here so Nano is cited as documentation, not marketing.

  • Text inside images

    Flux.1 renders short captions reasonably well but degrades on paragraphs or stylised typography. Add text in post (Photoshop / Figma) for production-quality typography.

  • Sub-12 GB VRAM throughput

    The streaming DiT path runs the same weights on 8 GB / 10 GB cards but at roughly 0.4-0.6× throughput. Output quality is identical; only wall-clock differs.

  • Hands and fine anatomy

    Like other open-weight diffusion models, Flux.1 still produces occasional finger-count and joint errors. Use img2img inpaint for hand fixes, or generate hands separately and composite.

  • Identity preservation across edits

    Image-to-image with high strength may drift the subject's identity. For face-stable workflows, route the result through Nano FacialEdit's identity-preserving stack.

  • Apple Silicon is not supported in v1.0.5

    Current build is Windows + NVIDIA CUDA only. Metal support is on the roadmap.

  • Output licensing follows Flux.1 model license

    Generated images are usable commercially under the open Flux.1 license plus the NanoPocket Terms of Use. Subjects and trademarks must still be cleared by the user.

Methodology

Generation throughput is reported in seconds per image at 1024 × 1024, 30 sampling steps, fp16, batch=1. Tested on RTX 4070 Ti (12 GB VRAM), RTX 4080 (16 GB VRAM), and RTX 4090 (24 GB VRAM) on Windows 11 23H2 driver 553.62. Prompt-following uses GenEval and DrawBench public protocols on internal 200-prompt subsets.

External references

Frequently asked questions

7 buyer-voice questions about Nano, answered by the team.

Will Flux.1 actually run on my GPU?+

If the GPU has at least 12 GB of VRAM (RTX 3060 / 4060 / 5060 class or better), yes. Below that, the streaming DiT path still loads the model in chunks; smaller cards generate at lower resolution and lower throughput. Cards with more VRAM run the same model faster.

How is this different from Midjourney, DALL-E, or Adobe Firefly?+

Midjourney, DALL-E (OpenAI), Adobe Firefly, Leonardo.Ai, Krea, and ideogram run server-side and charge per generation or per seat. Nano ImageEdit runs the open-weight Flux.1 model on the user's local GPU with a one-time license, no per-image charge, and no remote prompt logging.

Can I edit an existing photo, not just generate from text?+

Yes. Image-to-image accepts a reference image plus a prompt and applies the edit while preserving the reference's structural layout. A per-image strength slider controls how aggressively the prompt overrides the reference. Reference-image conditioning is also available for style transfer.

Do I need an internet connection to generate?+

No. After the one-time license-activation handshake, generation runs fully offline on the local GPU. Prompts, references, and outputs are not transmitted off the machine.

Will my prompts stay private?+

Yes. There is no remote endpoint contact during generation; the only network call is product-bound license activation. Prompts, references, and outputs stay on the user's disk and are not used for any model training.

Does it work on a Mac?+

Not in v1.0.5. The current build is Windows 10/11 + NVIDIA CUDA. Apple Silicon support is on the roadmap.

Can I use the generated images commercially?+

Yes. The license is one-time and machine-bound, with no per-image fees. Generated images can be used in commercial deliverables under the standard NanoPocket Terms of Use and the open Flux.1 model license. The user retains full output rights.

Install Nano ImageEdit.

Flux.1 image generation on your local GPU, with text-to-image, image-to-image, and a streaming DiT path for under-12 GB VRAM cards. No subscription, no cloud generation.